test: 新增统一回归运行器

新增 run-regressions.mjs 一键执行全部回归脚本，package.json 注册 npm test 入口与 test:fine-tune-create，page-surface 脚本适配统一运行。

frontend/package.json

@@ -8,6 +8,7 @@
     "build": "vue-tsc -b && vite build",
     "preview": "vite preview",
     "type-check": "vue-tsc -b --noEmit",
+    "test": "node scripts/run-regressions.mjs",
     "test:default-dashboard": "node scripts/regression-default-dashboard.mjs",
     "test:dashboard": "node scripts/regression-dashboard.mjs",
     "test:data-process-list": "node scripts/regression-data-process-list.mjs",
@@ -21,6 +22,7 @@
     "test:model-manage": "node scripts/regression-model-manage.mjs",
     "test:hardware": "node scripts/regression-hardware-dashboard.mjs",
     "test:training-log-layout": "node scripts/regression-training-log-layout.mjs",
+    "test:fine-tune-create": "node scripts/regression-fine-tune-create-ui.mjs",
     "test:page-surface": "node scripts/regression-page-surface.mjs"
   },
   "dependencies": {

frontend/scripts/regression-data-process-wizard.mjs

@@ -13,6 +13,13 @@ const confirmDialogPath = path.resolve(scriptDir, '../src/components/AppConfirmD
 const layoutPath = path.resolve(scriptDir, '../src/layouts/MainLayout.vue')
 const viewSource = await readFile(viewPath, 'utf8')
 const layoutSource = await readFile(layoutPath, 'utf8')
+const [draftSource, stateSource, generationSource, viewStyleSource] = await Promise.all([
+  readFile(path.join(createDir, 'useDataProcessDraft.ts'), 'utf8'),
+  readFile(path.join(createDir, 'dataProcessCreateState.ts'), 'utf8'),
+  readFile(path.join(createDir, 'useDataProcessGeneration.ts'), 'utf8'),
+  readFile(path.join(createDir, 'data-process-create.scss'), 'utf8'),
+])
+const implementationSource = [viewSource, draftSource, stateSource, generationSource].join('\n')
 
 assert.ok(existsSync(confirmDialogPath), '缺少公共确认弹窗组件 AppConfirmDialog')
 const confirmDialogSource = await readFile(confirmDialogPath, 'utf8')
@@ -43,9 +50,10 @@ assert.match(
   '五步向导顺序必须为创建任务、上传文件、数据预览、开始生成、结果编辑与保存',
 )
 assert.doesNotMatch(viewSource, /steps\s*=\s*computed|all\.filter/, '步骤仍根据处理类型动态增减')
-assert.match(viewSource, /localStorage\.setItem\(DRAFT_STORAGE_KEY/, '草稿没有持久化')
+assert.match(draftSource, /localStorage\.setItem\(DATA_PROCESS_DRAFT_STORAGE_KEY/, '草稿没有持久化')
-assert.match(viewSource, /localStorage\.getItem\(DRAFT_STORAGE_KEY\)/, '草稿没有恢复读取')
+assert.match(draftSource, /localStorage\.getItem\(DATA_PROCESS_DRAFT_STORAGE_KEY\)/, '草稿没有恢复读取')
 assert.match(viewSource, /restoreDraft\(\)/, '页面没有恢复草稿')
+assert.ok(viewSource.split('\n').length < 800, 'DataProcessCreateView 拆分后仍超过 800 行')
 
 const expectedComponents = [
   'TaskSetupStep.vue',
@@ -125,19 +133,49 @@ assert.match(previewSource, /\.preview-editor\s*\{[\s\S]*?flex:\s*1 1 auto[\s\S]
 assert.match(previewSource, /@media \(max-width: 900px\)/, '第三步缺少窄屏上下布局')
 
 const taskSetupPath = path.join(createDir, 'TaskSetupStep.vue')
-const taskSetupSource = await readFile(taskSetupPath, 'utf8')
+const structuredOptionsPath = path.join(createDir, 'StructuredOptionsPanel.vue')
+const unstructuredOptionsPath = path.join(createDir, 'UnstructuredOptionsPanel.vue')
+const datasetSplitEditorPath = path.join(createDir, 'DatasetSplitEditor.vue')
 const generationOptionsPath = path.join(createDir, 'GenerationOptionsPanel.vue')
-const generationControlSource = await readFile(generationOptionsPath, 'utf8')
 const sourceUploadPath = path.join(createDir, 'SourceUploadStep.vue')
-const sourceUploadSource = await readFile(sourceUploadPath, 'utf8')
+const [
+  taskSetupSource,
+  structuredOptionsSource,
+  unstructuredOptionsSource,
+  datasetSplitEditorSource,
+  generationControlSource,
+  sourceUploadSource,
+] = await Promise.all([
+  readFile(taskSetupPath, 'utf8'),
+  readFile(structuredOptionsPath, 'utf8'),
+  readFile(unstructuredOptionsPath, 'utf8'),
+  readFile(datasetSplitEditorPath, 'utf8'),
+  readFile(generationOptionsPath, 'utf8'),
+  readFile(sourceUploadPath, 'utf8'),
+])
+const taskSetupFeatureSource = [
+  taskSetupSource,
+  structuredOptionsSource,
+  unstructuredOptionsSource,
+  datasetSplitEditorSource,
+].join('\n')
+
+for (const componentPath of [structuredOptionsPath, unstructuredOptionsPath, datasetSplitEditorPath]) {
+  assert.ok(existsSync(componentPath), `缺少任务配置拆分组件：${path.basename(componentPath)}`)
+}
+assert.ok(taskSetupSource.split('\n').length < 800, 'TaskSetupStep 拆分后仍超过 800 行')
+assert.match(taskSetupSource, /<StructuredOptionsPanel/, '任务配置没有挂载结构化选项面板')
+assert.match(taskSetupSource, /<UnstructuredOptionsPanel/, '任务配置没有挂载非结构化选项面板')
+assert.match(structuredOptionsSource, /<DatasetSplitEditor/, '结构化选项没有复用数据集划分编辑器')
+assert.match(unstructuredOptionsSource, /<DatasetSplitEditor/, '非结构化选项没有复用数据集划分编辑器')
 
 for (const marker of ['<el-upload', '源数据上传', '数据源配置', 'uploadedFiles']) {
-  assert.ok(!taskSetupSource.includes(marker), `第一步仍包含上传职责：${marker}`)
+  assert.ok(!taskSetupFeatureSource.includes(marker), `第一步仍包含上传职责：${marker}`)
 }
 assert.match(viewSource, /<SourceUploadStep\s+[\s\S]*?v-else-if="currentStepId === 'upload'"/, '第二步没有挂载独立上传组件')
 assert.match(viewSource, /if \(currentStepId\.value === 'create'\) return '继续：上传文件'/, '第一步主按钮没有指向上传文件')
 assert.match(viewSource, /if \(currentStepId\.value === 'upload'\) return '继续：数据预览'/, '第二步主按钮没有指向数据预览')
-assert.match(viewSource, /const DRAFT_SCHEMA_VERSION = 5/, '生成配置扩展后必须升级草稿版本')
+assert.match(draftSource, /DATA_PROCESS_DRAFT_SCHEMA_VERSION = 6/, '安全草稿格式必须升级到 v6')
 
 const nextFromCreateStart = viewSource.indexOf('async function nextFromCreate()')
 const nextFromUploadStart = viewSource.indexOf('function nextFromUpload()', nextFromCreateStart)
@@ -154,9 +192,15 @@ assert.match(nextFromUploadSource, /goToStep\('preview'\)/, '上传步骤完成
 assert.match(viewSource, /function goToStep\(stepId: StepId\)[\s\S]*?WIZARD_STEPS\.findIndex/, '向导跳转没有使用稳定步骤标识')
 assert.match(viewSource, /currentStepId\.value === 'preview'[\s\S]*?goToStep\('generate'\)/, '数据预览步骤没有进入开始生成')
 assert.match(viewSource, /generation\.status === 'success'[\s\S]*?goToStep\('results'\)/, '生成成功后没有进入结果编辑与保存')
-assert.match(viewSource, /currentStepId:\s*currentStepId\.value/, '草稿没有保存稳定步骤标识')
+assert.match(draftSource, /currentStepId:\s*bindings\.currentStepId\.value/, '草稿没有保存稳定步骤标识')
-assert.match(viewSource, /const LEGACY_V3_STEP_IDS:[^=]*= \['create', 'preview', 'generate', 'results'\]/, 'v3 草稿缺少旧步骤索引映射')
+const draftSnapshotSource = draftSource.slice(
-assert.match(viewSource, /snapshotSchemaVersion === 3[\s\S]*?LEGACY_V3_STEP_IDS/, 'v3 草稿没有迁移到五步索引')
+  draftSource.indexOf('function draftSnapshot()'),
+  draftSource.indexOf('function writeDraft'),
+)
+for (const forbiddenField of ['uploadedFiles', 'previewItems', 'results', 'password', 'token']) {
+  assert.ok(!draftSnapshotSource.includes(forbiddenField), `安全草稿不应持久化：${forbiddenField}`)
+}
+assert.match(draftSource, /writeDraft\(false\)/, '恢复旧草稿后没有立即覆盖潜在敏感数据')
 assert.match(viewSource, /<ResultEditorStep\s+[\s\S]*?v-else-if="currentStepId === 'results'"/, '结果编辑器必须只在结果步骤渲染')
 assert.match(viewSource, /watch\(processType,[\s\S]*?resetSourceDataForProcessTypeChange\(\)/, '切换处理类型后没有失效旧源数据')
 assert.match(viewSource, /function resetSourceDataForProcessTypeChange\(\)[\s\S]*?uploadedFiles\.value = \[\][\s\S]*?selectedPreviewFileId\.value = null/, '旧源数据失效没有同步清理文件与预览选择')
@@ -170,32 +214,32 @@ for (const option of [
   '异常数据过滤',
   '敏感信息脱敏',
 ]) {
-  assert.ok(taskSetupSource.includes(option), `结构化预处理缺少选项：${option}`)
+  assert.ok(structuredOptionsSource.includes(option), `结构化预处理缺少选项：${option}`)
 }
-assert.ok(taskSetupSource.includes('生成选项'), '结构化配置缺少生成选项分类')
+assert.ok(structuredOptionsSource.includes('生成选项'), '结构化配置缺少生成选项分类')
-assert.ok(taskSetupSource.includes('语义丰富表达'), '生成选项缺少语义丰富表达开关')
+assert.ok(structuredOptionsSource.includes('语义丰富表达'), '生成选项缺少语义丰富表达开关')
-assert.ok(taskSetupSource.includes('使用大模型将问答表述得更自然、柔和'), '语义丰富表达缺少辅助说明')
+assert.ok(structuredOptionsSource.includes('使用大模型将问答表述得更自然、柔和'), '语义丰富表达缺少辅助说明')
 for (const splitName of ['训练集', '验证集', '测试集']) {
-  assert.ok(taskSetupSource.includes(splitName), `生成选项缺少数据集划分：${splitName}`)
+  assert.ok(datasetSplitEditorSource.includes(splitName), `生成选项缺少数据集划分：${splitName}`)
 }
-assert.match(taskSetupSource, /const splitTotal = computed/, '数据集划分缺少比例总和计算')
+assert.match(datasetSplitEditorSource, /const splitTotal = computed/, '数据集划分缺少比例总和计算')
 assert.match(taskSetupSource, /splitTotal\.value !== 100/, '数据集划分缺少总和 100% 校验')
-assert.ok(taskSetupSource.includes('训练集、验证集和测试集比例总和必须为 100%'), '数据集划分缺少就地错误提示')
+assert.ok(datasetSplitEditorSource.includes('训练集、验证集和测试集比例总和必须为 100%'), '数据集划分缺少就地错误提示')
 for (const splitField of ['train', 'validation', 'test']) {
   assert.match(
-    taskSetupSource,
+    datasetSplitEditorSource,
-    new RegExp(`structuredOptions\\.datasetSplit\\.${splitField}[\\s\\S]*?:min="0"[\\s\\S]*?:max="100"[\\s\\S]*?:step="1"[\\s\\S]*?:precision="0"`),
+    new RegExp(`modelValue\\.${splitField}[\\s\\S]*?:min="0"[\\s\\S]*?:max="100"[\\s\\S]*?:step="1"[\\s\\S]*?:precision="0"`),
     `数据集划分字段 ${splitField} 缺少 0～100 的整数限制`,
   )
 }
-assert.match(taskSetupSource, /<el-switch[\s\S]*structuredOptions\.semanticEnrichment/, '语义丰富表达必须使用开关控件')
+assert.match(structuredOptionsSource, /<el-switch[\s\S]*options\.semanticEnrichment/, '语义丰富表达必须使用开关控件')
-assert.match(taskSetupSource, /<el-input-number[\s\S]*structuredOptions\.qaPairsPerRow[\s\S]*:min="1"[\s\S]*:max="5"/, '每行生成数量必须限制在 1 到 5')
+assert.match(structuredOptionsSource, /<el-input-number[\s\S]*options\.qaPairsPerRow[\s\S]*:min="1"[\s\S]*:max="5"/, '每行生成数量必须限制在 1 到 5')
 assert.match(viewSource, /const structuredOptions = ref<StructuredProcessOptions>/, '父页面缺少结构化配置状态')
-assert.match(viewSource, /datasetSplit:\s*\{ train: 80, validation: 10, test: 10 \}/, '数据集划分默认值必须为 80/10/10')
+assert.match(stateSource, /datasetSplit:\s*\{ train: 80, validation: 10, test: 10 \}/, '数据集划分默认值必须为 80/10/10')
-assert.match(viewSource, /structuredOptions:\s*\{[\s\S]*\.\.\.structuredOptions\.value/, '结构化配置没有写入草稿')
+assert.match(draftSource, /structuredOptions:\s*\{[\s\S]*\.\.\.bindings\.structuredOptions\.value/, '结构化配置没有写入草稿')
-assert.match(viewSource, /structuredOptions\.value = \{[\s\S]*\.\.\.snapshot\.structuredOptions/, '结构化配置没有从草稿恢复')
+assert.match(draftSource, /bindings\.structuredOptions\.value = \{[\s\S]*\.\.\.snapshot\.structuredOptions/, '结构化配置没有从草稿恢复')
 assert.match(viewSource, /v-model:structured-options="structuredOptions"/, '父页面没有双向绑定结构化配置')
-assert.match(viewSource, /createResults\([\s\S]*structuredOptions\.value/, '每行生成数量没有接入结果生成逻辑')
+assert.match(generationSource, /createResults\([\s\S]*bindings\.structuredOptions\.value/, '每行生成数量没有接入结果生成逻辑')
 
 for (const field of [
   'generationModelId',
@@ -206,23 +250,25 @@ for (const field of [
   'minOutputLength',
 ]) {
   assert.ok(typesSource.includes(field), `生成控制配置缺少字段：${field}`)
-  assert.ok(viewSource.includes(field), `父页面默认值或草稿状态缺少字段：${field}`)
+  assert.ok(implementationSource.includes(field), `父页面默认值或草稿状态缺少字段：${field}`)
 }
-assert.match(taskSetupSource, /GenerationOptionsPanel/, '生成选项没有复用统一的大模型与质量筛选组件')
+assert.match(structuredOptionsSource, /GenerationOptionsPanel/, '结构化生成选项没有复用统一的大模型与质量筛选组件')
-assert.match(taskSetupSource, /:options="structuredOptions"/, '结构化生成选项未接入统一配置组件')
+assert.match(unstructuredOptionsSource, /GenerationOptionsPanel/, '非结构化生成选项没有复用统一的大模型与质量筛选组件')
-assert.match(taskSetupSource, /:options="unstructuredOptions"/, '非结构化生成选项未接入统一配置组件')
+assert.match(structuredOptionsSource, /:options="options"/, '结构化生成选项未接入统一配置组件')
-assert.match(taskSetupSource, /<h3>大模型<\/h3>/, '大模型必须作为与生成选项平级的独立分类')
+assert.match(unstructuredOptionsSource, /:options="options"/, '非结构化生成选项未接入统一配置组件')
-assert.match(taskSetupSource, /section="model"/, '独立大模型分类没有挂载模型配置')
+assert.match(taskSetupFeatureSource, /<h3>大模型<\/h3>/, '大模型必须作为与生成选项平级的独立分类')
-assert.match(taskSetupSource, /section="quality"/, '质量筛选没有保留在生成选项分类中')
+assert.match(taskSetupFeatureSource, /section="model"/, '独立大模型分类没有挂载模型配置')
+assert.match(taskSetupFeatureSource, /section="quality"/, '质量筛选没有保留在生成选项分类中')
 assert.doesNotMatch(generationControlSource, /<h4>大模型<\/h4>/, '大模型不应继续作为生成选项内部子分类')
 for (const label of ['大模型', '数据生成模型', '默认提示语', '质量筛选', '过滤低质量内容', '过滤过短内容', '最少字数']) {
   assert.ok(generationControlSource.includes(label), `生成控制界面缺少：${label}`)
 }
 assert.match(generationControlSource, /filterable/, '数据生成模型下拉必须支持搜索')
 assert.match(generationControlSource, /maxlength="500"/, '默认提示语缺少合理的长度限制')
-assert.match(generationControlSource, /\.model-option-row\s*\{[\s\S]*?display:\s*flex[\s\S]*?border:\s*1px solid #e2e5ec/, '大模型配置没有与上方生成选项使用一致的配置行样式')
+assert.match(generationControlSource, /\.model-field\s*\{[\s\S]*?display:\s*flex[\s\S]*?flex-direction:\s*column/, '大模型字段没有使用稳定的纵向表单布局')
-assert.match(viewSource, /const DEFAULT_GENERATION_PROMPT\s*=\s*['"][^'"]{40,}['"]/, '大模型配置缺少可直接使用的默认提示语')
+assert.match(generationControlSource, /\.generation-config-group\s*\{[\s\S]*?border:\s*1px solid #e2e5ec/, '大模型配置没有保留统一配置面板边框')
-assert.equal((viewSource.match(/generationPrompt:\s*DEFAULT_GENERATION_PROMPT/g) || []).length, 2, '结构化与非结构化任务必须共用默认提示语')
+assert.match(stateSource, /const DEFAULT_GENERATION_PROMPT\s*=\s*['"][^'"]{40,}['"]/, '大模型配置缺少可直接使用的默认提示语')
+assert.equal((stateSource.match(/generationPrompt:\s*DEFAULT_GENERATION_PROMPT/g) || []).length, 2, '结构化与非结构化任务必须共用默认提示语')
 assert.match(generationControlSource, /v-if="options\.qualityFilterEnabled"/, '质量规则没有随总开关渐进显示')
 assert.match(generationControlSource, /v-if="options\.filterShortContent"/, '最少字数没有随短内容规则显示')
 assert.match(generationControlSource, /:min="1"[\s\S]*:max="1000"/, '最少字数缺少 1 到 1000 的边界限制')
@@ -253,60 +299,55 @@ for (const removedField of ['contextScope', 'generationTypes', 'skipUnanswerable
 }
 
 assert.match(taskSetupSource, /v-if="processType === 'unstructured'"/, '非结构化配置必须仅在非结构化数据类型下显示')
-assert.ok(taskSetupSource.includes('智能预处理'), '简化后缺少智能预处理总开关')
+assert.ok(unstructuredOptionsSource.includes('智能预处理'), '简化后缺少智能预处理总开关')
-assert.ok(taskSetupSource.includes('敏感信息脱敏'), '简化后缺少脱敏开关')
+assert.ok(unstructuredOptionsSource.includes('敏感信息脱敏'), '简化后缺少脱敏开关')
-assert.match(taskSetupSource, /const smartPreprocessEnabled = computed/, '智能预处理没有映射到内部处理项')
+assert.match(unstructuredOptionsSource, /const smartPreprocessEnabled = computed/, '智能预处理没有映射到内部处理项')
-assert.match(taskSetupSource, /function updateSmartPreprocess/, '智能预处理开关缺少更新逻辑')
+assert.match(unstructuredOptionsSource, /function updateSmartPreprocess/, '智能预处理开关缺少更新逻辑')
-assert.match(taskSetupSource, /function updateDesensitize/, '脱敏开关缺少更新逻辑')
+assert.match(unstructuredOptionsSource, /function updateDesensitize/, '脱敏开关缺少更新逻辑')
 
-assert.ok(taskSetupSource.includes('切分选项'), '非结构化配置缺少切分选项分类')
+assert.ok(unstructuredOptionsSource.includes('切分选项'), '非结构化配置缺少切分选项分类')
 for (const method of ['自动语义切分', '按标题和段落', '按固定长度', '自定义分隔符']) {
-  assert.ok(taskSetupSource.includes(method), `切分方式缺少选项：${method}`)
+  assert.ok(unstructuredOptionsSource.includes(method), `切分方式缺少选项：${method}`)
 }
 for (const label of ['切片长度', '重叠长度', '最小切片长度', '保护表格、代码和列表']) {
-  assert.ok(taskSetupSource.includes(label), `切分选项缺少配置：${label}`)
+  assert.ok(unstructuredOptionsSource.includes(label), `切分选项缺少配置：${label}`)
 }
-assert.ok(taskSetupSource.includes('高级设置'), '切分选项缺少渐进式高级设置入口')
+assert.ok(unstructuredOptionsSource.includes('高级设置'), '切分选项缺少渐进式高级设置入口')
-assert.match(taskSetupSource, /:aria-expanded="advancedChunkSettingsOpen"/, '高级设置入口缺少展开状态的可访问性标记')
+assert.match(unstructuredOptionsSource, /:aria-expanded="advancedChunkSettingsOpen"/, '高级设置入口缺少展开状态的可访问性标记')
-assert.match(taskSetupSource, /v-if="advancedChunkSettingsOpen"/, '高级设置内容没有默认收起')
+assert.match(unstructuredOptionsSource, /v-if="advancedChunkSettingsOpen"/, '高级设置内容没有默认收起')
-assert.match(taskSetupSource, /watch\(\(\) => props\.unstructuredOptions\.chunkMethod,[\s\S]*?method === 'custom'[\s\S]*?advancedChunkSettingsOpen\.value = true/, '选择自定义分隔符时没有自动展开高级设置')
+assert.match(unstructuredOptionsSource, /watch\(\(\) => props\.options\.chunkMethod,[\s\S]*?method === 'custom'[\s\S]*?advancedChunkSettingsOpen\.value = true/, '选择自定义分隔符时没有自动展开高级设置')
-assert.match(taskSetupSource, /if \(chunkValidationMessage\.value\) \{[\s\S]*?advancedChunkSettingsOpen\.value = true[\s\S]*?return false/, '高级切分配置校验失败时没有重新展开定位')
+assert.match(unstructuredOptionsSource, /function revealValidation\(\)[\s\S]*?advancedChunkSettingsOpen\.value = true/, '非结构化面板没有暴露高级校验定位能力')
-assert.match(taskSetupSource, /const preserveSpecialContentEnabled = computed/, '特殊内容保护没有合并为单一开关')
+assert.match(taskSetupSource, /if \(chunkValidationMessage\.value\) \{[\s\S]*?revealValidation\(\)[\s\S]*?return false/, '高级切分配置校验失败时没有重新展开定位')
-assert.match(taskSetupSource, /function updateSpecialContentProtection/, '特殊内容保护开关缺少更新逻辑')
+assert.match(unstructuredOptionsSource, /const preserveSpecialContentEnabled = computed/, '特殊内容保护没有合并为单一开关')
-assert.match(taskSetupSource, /unstructuredOptions\.chunkSize[\s\S]*?:min="200"[\s\S]*?:max="2000"/, '切片长度必须限制在 200 到 2000 Token')
+assert.match(unstructuredOptionsSource, /function updateSpecialContentProtection/, '特殊内容保护开关缺少更新逻辑')
-assert.match(taskSetupSource, /unstructuredOptions\.chunkOverlap[\s\S]*?:min="0"[\s\S]*?:max="500"/, '重叠长度必须限制在 0 到 500 Token')
+assert.match(unstructuredOptionsSource, /options\.chunkSize[\s\S]*?:min="200"[\s\S]*?:max="2000"/, '切片长度必须限制在 200 到 2000 Token')
-assert.match(taskSetupSource, /unstructuredOptions\.minChunkSize[\s\S]*?:min="20"[\s\S]*?:max="500"/, '最小切片长度必须限制在 20 到 500 Token')
+assert.match(unstructuredOptionsSource, /options\.chunkOverlap[\s\S]*?:min="0"[\s\S]*?:max="500"/, '重叠长度必须限制在 0 到 500 Token')
+assert.match(unstructuredOptionsSource, /options\.minChunkSize[\s\S]*?:min="20"[\s\S]*?:max="500"/, '最小切片长度必须限制在 20 到 500 Token')
 
 for (const label of ['语义丰富表达', '每个切片生成数量', '数据集划分']) {
-  assert.ok(taskSetupSource.includes(label), `非结构化生成选项缺少：${label}`)
+  assert.ok(taskSetupFeatureSource.includes(label), `非结构化生成选项缺少：${label}`)
 }
 for (const removedLabel of ['上下文范围', '问题类型', '跳过无法回答的内容']) {
-  assert.ok(!taskSetupSource.includes(removedLabel), `简化后仍显示低频选项：${removedLabel}`)
+  assert.ok(!taskSetupFeatureSource.includes(removedLabel), `简化后仍显示低频选项：${removedLabel}`)
 }
-assert.match(taskSetupSource, /unstructuredOptions\.qaPairsPerChunk[\s\S]*?:min="1"[\s\S]*?:max="3"/, '每个切片生成数量必须限制在 1 到 3')
+assert.match(unstructuredOptionsSource, /options\.qaPairsPerChunk[\s\S]*?:min="1"[\s\S]*?:max="3"/, '每个切片生成数量必须限制在 1 到 3')
 assert.match(taskSetupSource, /unstructuredSplitTotal\.value !== 100/, '非结构化数据集划分缺少总和 100% 校验')
 assert.match(taskSetupSource, /chunkOverlap \+ props\.unstructuredOptions\.minChunkSize[\s\S]*?> props\.unstructuredOptions\.chunkSize/, '切分配置未校验重叠长度与最小切片长度的组合边界')
-assert.ok(taskSetupSource.includes('Token 数为轻量估算值'), '切片长度缺少 Token 估算说明')
+assert.ok(unstructuredOptionsSource.includes('Token 数为轻量估算值'), '切片长度缺少 Token 估算说明')
-assert.match(taskSetupSource, /\.chunk-settings-grid\s*\{[\s\S]*?grid-template-columns:\s*repeat\(3,\s*minmax\(0,\s*1fr\)\)/, '核心切分参数没有收紧为三列布局')
+assert.match(unstructuredOptionsSource, /\.chunk-settings-grid\s*\{[\s\S]*?grid-template-columns:\s*repeat\(3,\s*minmax\(0,\s*1fr\)\)/, '核心切分参数没有收紧为三列布局')
 
 assert.match(viewSource, /const unstructuredOptions = ref<UnstructuredProcessOptions>/, '父页面缺少非结构化配置状态')
-assert.match(viewSource, /chunkMethod:\s*'semantic'/, '非结构化默认切分方式必须为自动语义切分')
+assert.match(stateSource, /chunkMethod:\s*'semantic'/, '非结构化默认切分方式必须为自动语义切分')
-assert.match(viewSource, /chunkSize:\s*800/, '默认切片长度必须为 800 Token')
+assert.match(stateSource, /chunkSize:\s*800/, '默认切片长度必须为 800 Token')
-assert.match(viewSource, /chunkOverlap:\s*100/, '默认重叠长度必须为 100 Token')
+assert.match(stateSource, /chunkOverlap:\s*100/, '默认重叠长度必须为 100 Token')
-assert.match(viewSource, /minChunkSize:\s*100/, '默认最小切片长度必须为 100 Token')
+assert.match(stateSource, /minChunkSize:\s*100/, '默认最小切片长度必须为 100 Token')
-assert.match(viewSource, /qaPairsPerChunk:\s*1/, '默认每个切片必须生成 1 个问答对')
+assert.match(stateSource, /qaPairsPerChunk:\s*1/, '默认每个切片必须生成 1 个问答对')
-assert.match(viewSource, /unstructuredOptions:\s*\{[\s\S]*\.\.\.unstructuredOptions\.value/, '非结构化配置没有写入草稿')
+assert.match(draftSource, /unstructuredOptions:\s*\{[\s\S]*\.\.\.bindings\.unstructuredOptions\.value/, '非结构化配置没有写入草稿')
-assert.match(viewSource, /unstructuredOptions\.value = \{[\s\S]*restoredUnstructuredOptions\.chunkSize/, '非结构化配置没有从草稿恢复')
+assert.match(draftSource, /bindings\.unstructuredOptions\.value = \{[\s\S]*\.\.\.snapshot\.unstructuredOptions/, '非结构化配置没有从草稿恢复')
 assert.match(viewSource, /v-model:unstructured-options="unstructuredOptions"/, '父页面没有双向绑定非结构化配置')
-assert.match(viewSource, /const DRAFT_SCHEMA_VERSION = \d+/, '草稿缺少版本标识')
+assert.match(draftSource, /DATA_PROCESS_DRAFT_SCHEMA_VERSION = \d+/, '草稿缺少版本标识')
-assert.match(viewSource, /schemaVersion:\s*DRAFT_SCHEMA_VERSION/, '草稿快照没有写入版本标识')
+assert.match(draftSource, /schemaVersion:\s*DATA_PROCESS_DRAFT_SCHEMA_VERSION/, '草稿快照没有写入版本标识')
-assert.match(viewSource, /requiresPreviewMigration/, '旧草稿没有失效旧切片预览')
+assert.match(draftSource, /bindings\.goToStep\('create'\)/, '恢复配置后应回到安全的创建步骤')
-assert.doesNotMatch(
-  viewSource,
-  /snapshot\.unstructuredOptions\s*&&\s*!requiresPreviewMigration/,
-  '草稿版本迁移不应丢失仍受支持的非结构化配置',
-)
-assert.match(viewSource, /const restoredUnstructuredOptions = snapshot\.unstructuredOptions/, '旧草稿没有通过白名单恢复仍受支持的配置')
 assert.match(viewSource, /JSON\.stringify\(previewAffectingOptions\(\)\)/, '影响切分的非结构化配置没有纳入预览失效判断')
 
 const previewOptionsStart = viewSource.indexOf('function previewAffectingOptions()')
@@ -810,10 +851,10 @@ assert.match(sourceUploadSource, /@click="emit\('remove-file', file\.uid\)"/, '
 const { descriptor } = parseSfc(viewSource, { filename: viewPath })
 const template = descriptor.template?.content || ''
 assert.equal((template.match(/class="wizard-primary-action"/g) || []).length, 1, '页面必须只有一个主操作入口')
-assert.match(viewSource, /onBeforeUnmount\(stopGenerationTimer\)/, '生成计时器没有在卸载时清理')
+assert.match(viewSource, /onBeforeUnmount\(\(\) => \{[\s\S]*?stopGenerationTimer\(\)[\s\S]*?clearTimeout\(connectionTimer\)[\s\S]*?clearTimeout\(pullTimer\)/, '生成与外部数据源计时器没有在卸载时清理')
 assert.match(viewSource, /function scrollToStepTop/, '步骤切换后没有恢复页面顶部上下文')
 assert.match(viewSource, /nextTick\(scrollToStepTop\)/, '步骤切换没有触发页面滚动复位')
-assert.match(viewSource, /\.wizard-content\s*\{[\s\S]*min-height:\s*400px/, '第一步内容区必须保留足够高度以显示底部操作栏')
+assert.match(viewStyleSource, /\.wizard-content\s*\{[\s\S]*min-height:\s*400px/, '第一步内容区必须保留足够高度以显示底部操作栏')
 assert.match(
   layoutSource,
   /&:has\(\.create-wizard-layout\)\s*\{[\s\S]*?overflow-y:\s*hidden[\s\S]*?\.page-canvas\s*\{[\s\S]*?flex:\s*1 1 auto[\s\S]*?min-height:\s*0/,

frontend/scripts/regression-dataset-preview.mjs

@@ -4,12 +4,24 @@ import ts from 'typescript'
 
 const root = process.cwd()
 const previewPath = path.join(root, 'src/views/dataset/DatasetPreviewView.vue')
+const previewComponentsDir = path.join(root, 'src/views/dataset/preview')
+const previewComponentPaths = [
+  'DatasetVersionBar.vue',
+  'DatasetRecordTable.vue',
+  'DatasetRecordEditorDialog.vue',
+  'DatasetRawPreview.vue',
+].map((name) => path.join(previewComponentsDir, name))
 const apiPath = path.join(root, 'src/api/modules/dataset.ts')
 const adapterPath = path.join(root, 'src/mock/adapter.ts')
 const mockPath = path.join(root, 'src/mock/data.ts')
 const recordsPath = path.join(root, 'src/views/dataset/datasetRecords.ts')
 const versionsPath = path.join(root, 'src/mock/datasetVersions.ts')
 const source = fs.readFileSync(previewPath, 'utf8')
+const previewComponentSources = previewComponentPaths.map((componentPath) => {
+  expect(fs.existsSync(componentPath), `缺少数据集预览叶子组件：${path.basename(componentPath)}`)
+  return fs.readFileSync(componentPath, 'utf8')
+})
+const previewSurfaceSource = [source, ...previewComponentSources].join('\n')
 const apiSource = fs.readFileSync(apiPath, 'utf8')
 const adapterSource = fs.readFileSync(adapterPath, 'utf8')
 const mockSource = fs.readFileSync(mockPath, 'utf8')
@@ -35,33 +47,55 @@ function expect(condition, message) {
   if (!condition) throw new Error(message)
 }
 
+expect(source.split('\n').length < 800, 'DatasetPreviewView.vue 应拆分到 800 行以内')
+for (const componentName of ['DatasetVersionBar', 'DatasetRecordTable', 'DatasetRecordEditorDialog', 'DatasetRawPreview']) {
+  expect(source.includes(`import ${componentName}`), `主页面未导入叶子组件：${componentName}`)
+  expect(source.includes(`<${componentName}`), `主页面未挂载叶子组件：${componentName}`)
+}
 expect(source.includes('class="preview-workspace"'), '详情页应提供数据文件工作区')
 expect(!source.includes('class="file-pane"'), '单文件数据集详情不应展示文件选择侧栏')
-expect(source.includes('class="code-line"'), '内容查看器应提供逐行展示')
+expect(previewSurfaceSource.includes('class="code-line"'), '内容查看器应提供逐行展示')
-expect(source.includes('class="line-number"'), '内容查看器应显示行号')
+expect(previewSurfaceSource.includes('class="line-number"'), '内容查看器应显示行号')
 expect(source.includes('previewLoading'), '切换文件时应提供独立加载状态')
 expect(source.includes('previewError'), '内容加载失败时应提供错误状态')
-expect(source.includes('class="records-viewer"'), '详情页应将结构化文件展示为样本列表')
+expect(previewSurfaceSource.includes('class="records-viewer"'), '详情页应将结构化文件展示为样本列表')
-expect(source.includes('class="record-table"'), '样本数据应使用企业级表格展示')
+expect(previewSurfaceSource.includes('class="record-table"'), '样本数据应使用企业级表格展示')
-expect(source.includes('fixed="right"'), '逐条编辑操作列应固定在表格右侧')
+expect(previewSurfaceSource.includes('fixed="right"'), '逐条编辑操作列应固定在表格右侧')
-expect(source.includes('v-for="fieldKey in tableFieldKeys"'), '企业表格应根据数据结构动态生成字段列')
+expect(previewSurfaceSource.includes('v-for="fieldKey in tableFieldKeys"'), '企业表格应根据数据结构动态生成字段列')
-expect(source.includes('openRecordEditor(asDatasetRecord(row))'), '每条样本应提供独立编辑入口')
+expect(previewSurfaceSource.includes("emit('edit', asDatasetRecord(row))"), '每条样本应向主页面发送编辑事件')
-expect(source.includes('暂存修改'), '单条编辑应先暂存到当前页面')
+expect(previewSurfaceSource.includes('暂存修改'), '单条编辑应先暂存到当前页面')
 expect(source.includes('hasPendingVersionChanges && isViewingActiveVersion'), '存在暂存修改时才应显示保存版本按钮')
 expect(source.includes('保存版本'), '文件工具栏应提供保存版本按钮')
 expect(source.includes('baseVersionContent'), '页面应区分已保存版本与待保存工作副本')
-expect(source.includes('class="version-control-bar"'), '详情页应提供独立版本控制区')
+expect(previewSurfaceSource.includes('class="version-control-bar"'), '详情页应提供独立版本控制区')
-expect(source.includes('设为当前版本'), '历史版本应支持显式切换为当前版本')
+expect(previewSurfaceSource.includes('设为当前版本'), '历史版本应支持显式切换为当前版本')
-expect(source.includes('历史版本（只读）'), '历史版本应明确展示只读状态')
+expect(previewSurfaceSource.includes('历史版本（只读）'), '历史版本应明确展示只读状态')
 expect(source.includes('loadedVersionId.value'), '下载和展示应绑定正在查看的具体版本')
 expect(source.includes('versionRequestId'), '快速切换历史版本时应防止旧响应覆盖新内容')
-expect(source.includes('v-model:current-page="currentPage"'), '样本列表应支持分页浏览')
+expect(previewSurfaceSource.includes('v-model:current-page="currentPage"'), '样本列表应支持分页浏览')
 expect(!source.includes('class="content-editor"'), '详情页不应继续提供整文件编辑器')
 expect(source.includes('saveRecord'), '逐条编辑器应提供单条保存动作')
 expect(source.includes('hasUnsavedChanges'), '在线编辑器应跟踪未保存修改')
 expect(source.includes('onBeforeRouteLeave'), '离开页面时应保护未保存修改')
-expect(source.includes("event.key.toLowerCase() === 's'"), '在线编辑器应支持快捷键保存')
+expect(previewSurfaceSource.includes("event.key.toLowerCase() === 's'"), '在线编辑器应支持快捷键保存')
 expect(source.includes('loadVersions(selectedFile)'), '版本内容加载失败后应支持重新加载')
+expect(source.includes('function resetRecordEditor()'), '主页面应提供统一的编辑器状态清理函数')
+const resetEditorStart = source.indexOf('function resetRecordEditor()')
+const updateFieldStart = source.indexOf('function updateEditField', resetEditorStart)
+const resetEditorSource = source.slice(resetEditorStart, updateFieldStart)
+for (const marker of [
+  'editorVisible.value = false',
+  'editingRecord.value = null',
+  'editFields.value = []',
+  "rawDraft.value = ''",
+  "originalDraft.value = ''",
+]) {
+  expect(resetEditorSource.includes(marker), `编辑器状态清理不完整：${marker}`)
+}
+const versionChangeStart = source.indexOf('async function handleViewedVersionChange')
+const activateVersionStart = source.indexOf('async function activateViewedVersion', versionChangeStart)
+const versionChangeSource = source.slice(versionChangeStart, activateVersionStart)
+expect(versionChangeSource.includes('resetRecordEditor()'), '成功切换版本后必须关闭并清空旧编辑器状态')
 expect(!source.includes('handleDownloadAll'), '页面不应保留整包下载逻辑')
 expect(!source.includes('handleDelete'), '页面不应保留删除逻辑')
 expect(!source.includes('router.back()'), '页面不应保留页头返回逻辑')

frontend/scripts/regression-fine-tune-create-ui.mjs

@@ -1,10 +1,12 @@
 import { readFileSync } from 'node:fs'
 import { resolve } from 'node:path'
 import { fileURLToPath } from 'node:url'
+import ts from 'typescript'
 
 const root = resolve(fileURLToPath(new URL('..', import.meta.url)))
 const source = readFileSync(resolve(root, 'src/views/fine-tune/FineTuneCreateView.vue'), 'utf8')
 const modelDialogSource = readFileSync(resolve(root, 'src/components/ModelSelectDialog.vue'), 'utf8')
+const formModelSource = readFileSync(resolve(root, 'src/views/fine-tune/fineTuneFormModel.ts'), 'utf8')
 
 function assert(condition, message) {
   if (!condition) {
@@ -24,11 +26,32 @@ assert(modelDialogSource.includes('width="860px"'), 'Model dialog should use a c
 assert(modelDialogSource.includes('model-series-list'), 'Model dialog should include a model series list column')
 assert(modelDialogSource.includes('model-version-list'), 'Model dialog should include a snapshot/version list column')
 assert(modelDialogSource.includes('handleConfirm'), 'Model dialog should confirm the selected model before updating the form')
-assert(source.includes("auto_merge: false"), 'Auto merge should default to disabled')
+assert(formModelSource.includes('auto_merge: false'), 'Auto merge should default to disabled')
 assert(source.includes('v-if="form.train_type === \'SFT\'"'), 'Merge model settings should only be visible for SFT')
 assert(source.includes('content-position="left">合并模型'), 'SFT form should include a merge model section below data configuration')
 assert(source.includes('v-model="form.auto_merge"'), 'Merge model section should provide an auto merge selector')
 assert(source.includes('label="自动合并权重并保存"'), 'Auto merge selector should use a clear visible label')
-assert((source.match(/auto_merge: form\.train_type === 'SFT' && form\.auto_merge/g) || []).length === 2, 'Auto merge should be sent for SFT when creating and starting the task')
+assert(formModelSource.includes("auto_merge: form.train_type === 'SFT' && form.auto_merge"), 'Auto merge should be normalized by the shared payload builder')
+assert(source.includes('const payload = buildFineTunePayload(form, selectedGpus.value)'), 'Create and start should share one normalized payload')
+assert(source.includes('startFineTune({ ...payload, task_id: taskId })'), 'Start request should reuse the normalized payload')
+assert(source.includes('Object.assign(form, DEFAULT_TRAINING_PARAMS)'), 'Reset should reuse the canonical defaults')
+assert(!source.includes('const taskData = {'), 'The duplicated create payload should be removed')
+assert(!source.includes('const createRes: any'), 'Create response should use the API return type')
+assert(!source.includes('(check as any).exists'), 'Name-check response should use its API return type')
+assert(source.includes('任务名校验失败'), 'Name-check failures should be visible and block submission')
+
+const runnableSource = ts.transpileModule(
+  formModelSource.replace("import type { FineTuneStartPayload, FineTuneTask } from '@/types'", ''),
+  { compilerOptions: { module: ts.ModuleKind.ESNext, target: ts.ScriptTarget.ES2022 } },
+).outputText
+const model = await import(`data:text/javascript;base64,${Buffer.from(runnableSource).toString('base64')}`)
+const defaults = model.createDefaultFineTuneForm()
+const customized = { ...defaults, train_type: 'DPO', train_method: 'full', auto_merge: true, quantization_bit: 4 }
+const payload = model.buildFineTunePayload(customized, [0, 2])
+assert(payload.auto_merge === false, 'Non-SFT tasks must never enable auto merge')
+assert(payload.quantization_bit === 0, 'Full fine-tuning must never send QLoRA quantization')
+assert(payload.gpus.join(',') === '0,2', 'Selected GPUs should be preserved in the shared payload')
+const resetDefaults = { ...model.DEFAULT_TRAINING_PARAMS }
+assert(resetDefaults.lora_alpha === defaults.lora_alpha, 'Reset and initial defaults must share LoRA values')
 
 console.log('fine-tune create UI regression checks passed')

frontend/scripts/regression-page-surface.mjs

@@ -6,11 +6,12 @@ import { parse as parseTemplate } from '@vue/compiler-dom'
 import { parse as parseSfc } from '@vue/compiler-sfc'
 
 const scriptDir = path.dirname(fileURLToPath(import.meta.url))
-const [globalStyles, routerSource, mainLayoutSource, trainingLogSource, fineTuneCreateSource] = await Promise.all([
+const [globalStyles, routerSource, mainLayoutSource, trainingLogSource, trainingOverviewSource, fineTuneCreateSource] = await Promise.all([
   readFile(path.resolve(scriptDir, '../src/styles/index.scss'), 'utf8'),
   readFile(path.resolve(scriptDir, '../src/router/index.ts'), 'utf8'),
   readFile(path.resolve(scriptDir, '../src/layouts/MainLayout.vue'), 'utf8'),
   readFile(path.resolve(scriptDir, '../src/views/system/TrainingLogView.vue'), 'utf8'),
+  readFile(path.resolve(scriptDir, '../src/views/system/training-log/TrainingTaskOverview.vue'), 'utf8'),
   readFile(path.resolve(scriptDir, '../src/views/fine-tune/FineTuneCreateView.vue'), 'utf8'),
 ])
 
@@ -70,8 +71,10 @@ const selfSurfaceRoutes = [
   'fine-tune',
   'model-eval',
   'model-inference',
+  'model-inference/chat/:id',
   'model-manage',
   'data-process',
+  'data-process/create',
   'dataset',
 ]
 for (const routePath of selfSurfaceRoutes) {
@@ -138,8 +141,8 @@ assert.doesNotMatch(
 const layoutContentBlock = extractCssBlock(mainLayoutStyle, '.layout-content')
 assert.match(layoutContentBlock, /background-color:\s*var\(--app-shell-bg\);/, '主内容区未使用灰色外层背景')
 
-const pageCanvasBlock = extractCssBlock(mainLayoutStyle, '.page-canvas')
+const pageCanvasBlock = extractCssBlock(mainLayoutStyle, '\n.page-canvas {')
-assert.match(pageCanvasBlock, /min-height:\s*100%;/, '白色页面画布没有铺满可用高度')
+assert.match(pageCanvasBlock, /flex:\s*1 0 auto;/, '白色页面画布没有铺满可用高度')
 assert.match(pageCanvasBlock, /padding:\s*24px;/, '白色页面画布缺少统一内容内边距')
 assert.match(pageCanvasBlock, /border-radius:\s*16px;/, '白色页面画布圆角与参考不一致')
 assert.match(pageCanvasBlock, /background-color:\s*var\(--app-page-bg\);/, '全局页面画布未使用白色背景')
@@ -174,7 +177,9 @@ assert.match(rootPageCardBlock, /background-color:\s*transparent;/, '页面根
 assert.match(rootPageCardBlock, /border-radius:\s*0\s*!important;/, '页面根卡片仍形成第二层圆角边界')
 assert.match(rootPageCardBlock, /box-shadow:\s*none\s*!important;/, '页面根卡片仍形成重复卡片层级')
 
-const trainingLogStyle = parseSfc(trainingLogSource).descriptor.styles.map((item) => item.content).join('\n')
+const trainingLogStyle = [trainingLogSource, trainingOverviewSource]
+  .flatMap((source) => parseSfc(source).descriptor.styles.map((item) => item.content))
+  .join('\n')
 const businessSurfaceBlock = extractCssBlock(trainingLogStyle, '.profile-section,\n.runtime-panel')
 assert.match(
   businessSurfaceBlock,

frontend/scripts/regression-training-log-layout.mjs

@@ -2,17 +2,35 @@ import assert from 'node:assert/strict'
 import { readFile } from 'node:fs/promises'
 import { fileURLToPath } from 'node:url'
 import path from 'node:path'
+import ts from 'typescript'
 import { parse as parseTemplate } from '@vue/compiler-dom'
 import { parse as parseSfc } from '@vue/compiler-sfc'
 
 const scriptDir = path.dirname(fileURLToPath(import.meta.url))
 const viewPath = path.resolve(scriptDir, '../src/views/system/TrainingLogView.vue')
-const source = await readFile(viewPath, 'utf8')
+const overviewPath = path.resolve(scriptDir, '../src/views/system/training-log/TrainingTaskOverview.vue')
+const modelPath = path.resolve(scriptDir, '../src/views/system/training-log/trainingLogModel.ts')
+const [source, overviewSource, modelSource] = await Promise.all([
+  readFile(viewPath, 'utf8'),
+  readFile(overviewPath, 'utf8'),
+  readFile(modelPath, 'utf8'),
+])
 const { descriptor } = parseSfc(source, { filename: viewPath })
-const template = descriptor.template?.content || ''
+const { descriptor: overviewDescriptor } = parseSfc(overviewSource, { filename: overviewPath })
-const style = descriptor.styles.map((item) => item.content).join('\n')
+const template = [descriptor.template?.content, overviewDescriptor.template?.content].filter(Boolean).join('\n')
+const viewStyle = descriptor.styles.map((item) => item.content).join('\n')
+const overviewStyle = overviewDescriptor.styles.map((item) => item.content).join('\n')
+const style = `${viewStyle}\n${overviewStyle}`
 const templateAst = parseTemplate(template)
 
+const modelModuleCode = ts.transpileModule(modelSource, {
+  compilerOptions: { module: ts.ModuleKind.ESNext, target: ts.ScriptTarget.ES2022 },
+}).outputText
+const {
+  parseTrainingLog,
+  resolveTrainingLogFile,
+} = await import(`data:text/javascript;base64,${Buffer.from(modelModuleCode).toString('base64')}`)
+
 function findElements(node, predicate, result = []) {
   if (node?.type === 1 && predicate(node)) result.push(node)
   for (const child of node?.children || []) findElements(child, predicate, result)
@@ -50,41 +68,71 @@ function extractCssBlock(css, marker) {
   assert.fail(`样式规则缺少右花括号：${marker}`)
 }
 
-function relativeLuminance(hex) {
+const logFiles = [
-  const channels = hex
+  { file: 'first_pid111.log', name: 'first-task', size: '1 KB', pid: 111 },
-    .replace('#', '')
+  { file: 'target_pid222.log', name: 'target-task', size: '1 KB', pid: 222 },
-    .match(/.{2}/g)
+]
-    .map((channel) => Number.parseInt(channel, 16) / 255)
+assert.equal(
-    .map((value) => (value <= 0.03928 ? value / 12.92 : ((value + 0.055) / 1.055) ** 2.4))
+  resolveTrainingLogFile(logFiles, { process_id: 222, name: 'target-task' })?.file,
-  return 0.2126 * channels[0] + 0.7152 * channels[1] + 0.0722 * channels[2]
+  'target_pid222.log',
-}
+  '训练日志必须按 task.process_id 精确匹配文件 pid，不能只判断两者是否存在',
+)
+assert.equal(
+  resolveTrainingLogFile(logFiles, { process_id: 999, name: 'target-task' })?.file,
+  'target_pid222.log',
+  'PID 无匹配时应按任务名回退选择日志',
+)
+assert.equal(
+  resolveTrainingLogFile(logFiles, { process_id: 999, name: 'missing-task' }),
+  undefined,
+  '任务无匹配日志时不得静默退回第一份日志',
+)
 
-function contrastRatio(foreground, background) {
+const parsed = parseTrainingLog([
-  const lighter = Math.max(relativeLuminance(foreground), relativeLuminance(background))
+  "INFO {'epoch': .5, 'learning_rate': 1.2E-5, 'grad_norm': -2.5e+0, 'loss': +3.25}",
-  const darker = Math.min(relativeLuminance(foreground), relativeLuminance(background))
+  '***** train metrics *****',
-  return (lighter + 0.05) / (darker + 0.05)
+  'train_runtime = 1.7852e3',
-}
+  "'train_loss': -3.42e-1",
+  'epoch: 1.0',
+  '***** train metrics end *****',
+].join('\n'))
+assert.deepEqual(parsed.metrics.loss, [3.25], '指标解析应允许字段乱序和带符号数值')
+assert.deepEqual(parsed.metrics.gradNorm, [-2.5], '梯度范数应支持科学计数法')
+assert.deepEqual(parsed.metrics.lr, [1.2e-5], '学习率应支持大小写科学计数法')
+assert.deepEqual(
+  parsed.summary,
+  { epoch: '1.0', trainLoss: '-3.42e-1', runtime: '1.7852e3' },
+  '训练汇总应同时支持等号、冒号、可选引号和科学计数法',
+)
+assert.deepEqual(
+  parseTrainingLog('普通日志，无训练指标').summary,
+  { epoch: '', trainLoss: '', runtime: '' },
+  '每次解析必须返回全新的空汇总，避免保留上一份日志的旧值',
+)
+assert.match(
+  source,
+  /const currentTask = await loadTask\(\)[\s\S]*?loadLog\(currentTask\)/,
+  '刷新流程必须先加载 task，再使用该 task 选择日志',
+)
+assert.match(source, /import \{ getSystemInfo \} from '@\/api\/modules\/system'/, '训练概览必须复用系统 GPU 监控数据源')
+assert.match(source, /Promise\.all\(\[datasetPromise, loadLog\(currentTask\), loadGpuStatus\(\)\]\)/, 'GPU 状态必须和训练日志一起刷新')
+assert.match(source, /if \(refreshInFlight\) return/, '轮询刷新必须阻止并发重叠')
+assert.match(source, /onUnmounted\([\s\S]*?clearInterval\(timer\)/, '组件卸载时必须清理轮询定时器')
 
 const overview = findElements(
   templateAst,
   (node) => staticAttribute(node, 'class')?.split(/\s+/).includes('overview-layout'),
 )
-assert.equal(overview.length, 1, '双栏任务档案容器必须且只能存在一个')
+assert.equal(overview.length, 1, '标准任务概况容器必须且只能存在一个')
-
+assert.doesNotMatch(overviewSource, /<aside\b/, '任务概况仍保留左右侧栏结构')
-for (const expectedClass of ['task-profile', 'dataset-profile', 'runtime-panel', 'parameter-groups']) {
+assert.doesNotMatch(overviewSource, /<i class="fa\b/, '任务概况仍包含过多装饰性图标')
-  const matched = findElements(
-    templateAst,
-    (node) => staticAttribute(node, 'class')?.split(/\s+/).includes(expectedClass),
-  )
-  assert.equal(matched.length, 1, `缺少或重复布局结构：${expectedClass}`)
-}
 
 const toggleButtons = findElements(
   templateAst,
-  (node) => node.tag === 'button'
+  (node) => node.tag === 'el-button'
     && staticAttribute(node, 'class')?.split(/\s+/).includes('params-toggle-button'),
 )
-assert.equal(toggleButtons.length, 1, '参数折叠必须使用唯一的原生 button')
+assert.equal(toggleButtons.length, 1, '参数折叠必须使用唯一的标准按钮')
 
 const toggleButton = toggleButtons[0]
 assert.equal(boundExpression(toggleButton, 'aria-expanded'), 'paramsExpanded', '折叠按钮未绑定 aria-expanded')
@@ -96,33 +144,80 @@ const controlledRegions = findElements(
 )
 assert.equal(controlledRegions.length, 1, 'aria-controls 指向的参数内容区域不存在或重复')
 
+const pageTitleIndex = source.indexOf('id="task-page-title"')
+const summaryCardIndex = source.indexOf('id="training-overview-title"')
+const taskOverviewIndex = source.indexOf('<TrainingTaskOverview')
+assert.notEqual(pageTitleIndex, -1, '页面缺少独立的训练任务标题')
+assert.ok(pageTitleIndex < summaryCardIndex, '训练任务标题必须出现在训练概览之前')
+assert.ok(summaryCardIndex < taskOverviewIndex, '训练概览必须出现在任务信息之前')
+assert.equal((source.match(/<h1\b/g) || []).length, 1, '训练详情页必须且只能有一个一级标题')
+assert.doesNotMatch(overviewSource, /<h1\b/, '任务信息卡片不得重复渲染页面一级标题')
+assert.ok(overviewSource.includes('title="任务信息"'), '任务详情卡片缺少明确的“任务信息”标题')
+
 const firstChartIndex = template.indexOf('<!-- 训练曲线 -->')
 assert.notEqual(firstChartIndex, -1, '未找到训练曲线边界，无法限定首屏检查范围')
-assert.equal(
+assert.ok(template.includes('training-progress-panel'), '训练概览缺少独立的主进度区域')
-  template.slice(0, firstChartIndex).includes('<el-descriptions'),
+assert.ok(template.includes('training-metric-grid'), '训练概览缺少次级训练指标摘要')
-  false,
+assert.ok(template.includes('gpu-device-list'), '训练概览缺少紧凑的 GPU 设备列表')
-  '任务概览、数据集和训练参数仍使用带表格感的 el-descriptions',
+assert.ok(template.includes('gpu-list-header'), 'GPU 设备列表缺少企业表格式列标题')
+assert.ok(template.includes('gpu-device-row'), 'GPU 设备列表缺少统一行结构')
+assert.doesNotMatch(template, /training-summary-strip|gpu-device-card/, '训练概览仍保留同级宫格或 GPU 嵌套卡片结构')
+assert.ok(template.includes('id="training-overview-title"'), '训练概览缺少可访问的卡片标题')
+assert.ok(
+  template.indexOf('id="training-overview-title"') < template.indexOf('<TrainingTaskOverview'),
+  '训练概览必须位于任务档案之前，成为页面第一块内容',
 )
+assert.ok(
+  template.indexOf('training-progress-panel') < template.indexOf('training-metric-grid')
+    && template.indexOf('training-metric-grid') < template.indexOf('gpu-device-list'),
+  '训练概览必须按主进度、次级指标、GPU 设备列表的顺序展示',
+)
+assert.ok(template.includes('id="gpu-monitor-title"'), '训练概览缺少 GPU 运行状态区域')
+assert.ok(template.includes('计算利用率'), 'GPU 运行状态缺少计算利用率')
+assert.ok(template.includes('显存占用'), 'GPU 运行状态缺少显存占用')
+assert.ok(template.includes('每 5 秒刷新'), 'GPU 运行状态缺少刷新频率说明')
+assert.ok(template.includes(':aria-label="`GPU ${item.index} 计算利用率'), 'GPU 强度条缺少可访问说明')
+assert.match(source, /const GPU_PREVIEW_LIMIT = 4/, '多 GPU 默认预览数量必须限制为 4 张')
+assert.match(source, /const visibleGpuItems = computed/, '多 GPU 缺少渐进披露列表计算')
+assert.match(source, /gpuRuntimePriority/, '折叠状态必须优先展示异常和运行中的 GPU')
+assert.ok(template.includes('v-for="item in visibleGpuItems"'), 'GPU 列表没有使用受控预览数据')
+assert.ok(template.includes('taskGpuItems.length > GPU_PREVIEW_LIMIT'), 'GPU 数量超过预览上限时没有展开入口')
+
+const gpuToggleButtons = findElements(
+  templateAst,
+  (node) => node.tag === 'el-button'
+    && staticAttribute(node, 'class')?.split(/\s+/).includes('gpu-toggle-button'),
+)
+assert.equal(gpuToggleButtons.length, 1, '多 GPU 展开控制必须且只能存在一个')
+assert.equal(boundExpression(gpuToggleButtons[0], 'aria-expanded'), 'gpuExpanded', 'GPU 展开按钮未绑定 aria-expanded')
+assert.equal(staticAttribute(gpuToggleButtons[0], 'aria-controls'), 'gpu-device-list', 'GPU 展开按钮缺少正确的 aria-controls')
+assert.ok(template.includes('title="训练曲线"'), '训练曲线没有使用标准 PageCard 标题')
+assert.ok(template.includes('title="训练日志"'), '训练日志没有使用标准 PageCard 标题')
+assert.ok(template.includes('每 5 秒刷新'), '训练监控区缺少自动刷新状态说明')
+assert.ok((template.match(/<el-descriptions\b/g) || []).length >= 3, '任务概况和训练参数没有使用标准详情表格')
+assert.equal((template.match(/class="chart-section"/g) || []).length, 3, '三组训练曲线没有按上下结构完整展示')
+assert.doesNotMatch(template, /<el-row\b|<el-col\b/, '训练曲线仍保留左右分栏布局')
+assert.doesNotMatch(template, /class="chart-card"/, '训练曲线仍存在卡片嵌套')
+assert.doesNotMatch(template, /summary-metric is-primary/, '训练概览仍保留大块装饰性主色背景')
+
+for (const selector of ['.summary-card', '.parameters-card', '.metrics-panel', '.log-card']) {
+  const cardBlock = extractCssBlock(style, selector)
+  assert.match(cardBlock, /border-radius:\s*8px/, `${selector} 未统一为 8px 企业卡片圆角`)
+  assert.match(cardBlock, /box-shadow:\s*none/, `${selector} 仍保留不统一的浮层阴影`)
+}
 
 assert.ok(template.includes("task?.output_model_name || '暂未生成'"), '输出模型缺失值文案不正确')
 assert.ok(template.includes("task?.batch_size ?? '未配置'"), '训练参数缺失值文案不正确')
 
-const media1100 = extractCssBlock(style, '@media (max-width: 1100px)')
+const overviewLayoutBlock = extractCssBlock(overviewStyle, '.overview-layout')
-const overviewAt1100 = extractCssBlock(media1100, '.overview-layout')
+assert.match(overviewLayoutBlock, /width:\s*100%/, '任务概况没有使用全宽单列布局')
-assert.match(overviewAt1100, /grid-template-columns:\s*minmax\(0,\s*1fr\)/, '1100px 断点未将双栏改为单栏')
+assert.doesNotMatch(overviewLayoutBlock, /grid-template-columns/, '任务档案仍保留左右分栏规则')
 
-const media700 = extractCssBlock(style, '@media (max-width: 700px)')
+const overviewMedia700 = extractCssBlock(overviewStyle, '@media (max-width: 700px)')
-for (const selector of ['.dataset-metrics', '.runtime-list', '.parameter-grid']) {
+assert.ok(overviewMedia700.includes('.task-descriptions'), '700px 断点缺少任务详情表单列规则')
-  assert.ok(media700.includes(selector), `700px 断点缺少单列规则：${selector}`)
+assert.match(overviewMedia700, /display:\s*block/, '移动端任务详情表没有转换为单列')
-}
+const viewMedia600 = extractCssBlock(viewStyle, '@media (max-width: 600px)')
-assert.match(media700, /grid-template-columns:\s*minmax\(0,\s*1fr\)/, '700px 断点未设置单列网格')
+assert.ok(viewMedia600.includes('.parameter-descriptions'), '600px 断点缺少训练参数表单列规则')
-
+assert.match(viewMedia600, /display:\s*block/, '移动端训练参数表没有转换为单列')
-const mutedBlock = extractCssBlock(style, '.is-muted')
-const mutedColor = mutedBlock.match(/color:\s*(#[0-9a-f]{6})/i)?.[1]
-assert.ok(mutedColor, '未配置文本缺少明确颜色')
-assert.ok(
-  contrastRatio(mutedColor, '#ffffff') >= 4.5,
-  `未配置文本颜色 ${mutedColor} 与白色背景对比度不足 4.5:1`,
-)
 
 console.log('训练日志详情布局回归检查通过')

frontend/scripts/regression-training-log-mock.mjs

@@ -0,0 +1,77 @@
+import assert from 'node:assert/strict'
+import { readFile } from 'node:fs/promises'
+import { fileURLToPath } from 'node:url'
+import path from 'node:path'
+
+const scriptDir = path.dirname(fileURLToPath(import.meta.url))
+const mockDataPath = path.resolve(scriptDir, '../src/mock/data.ts')
+const mockAdapterPath = path.resolve(scriptDir, '../src/mock/adapter.ts')
+const modelPath = path.resolve(scriptDir, '../src/views/system/training-log/trainingLogModel.ts')
+const [mockData, mockAdapter, modelSource] = await Promise.all([
+  readFile(mockDataPath, 'utf8'),
+  readFile(mockAdapterPath, 'utf8'),
+  readFile(modelPath, 'utf8'),
+])
+
+function objectFromMarker(source, marker) {
+  const markerIndex = source.indexOf(marker)
+  assert.notEqual(markerIndex, -1, `未找到 Mock 数据标记：${marker}`)
+  const openBrace = source.lastIndexOf('{', markerIndex)
+  assert.notEqual(openBrace, -1, `Mock 数据缺少对象起点：${marker}`)
+
+  let depth = 0
+  for (let index = openBrace; index < source.length; index += 1) {
+    if (source[index] === '{') depth += 1
+    if (source[index] === '}') depth -= 1
+    if (depth === 0) return source.slice(openBrace, index + 1)
+  }
+
+  assert.fail(`Mock 数据缺少对象终点：${marker}`)
+}
+
+const taskNames = [
+  'finance-sft-001',
+  'legal-sft-002',
+  'medical-cpt-001',
+  'service-dpo-001',
+  'finance-sft-002',
+  'general-sft-001',
+]
+const commonTrainingFields = [
+  'output_model_name',
+  'batch_size',
+  'learning_rate',
+  'n_epochs',
+  'save_steps',
+  'lr_scheduler_type',
+  'max_length',
+  'warmup_ratio',
+  'weight_decay',
+]
+
+for (const taskName of taskNames) {
+  const task = objectFromMarker(mockData, `name: '${taskName}'`)
+  for (const field of commonTrainingFields) {
+    assert.match(task, new RegExp(`\\b${field}\\s*:`), `${taskName} 缺少训练参数：${field}`)
+  }
+
+  if (/train_method:\s*'lora'/.test(task)) {
+    for (const field of ['lora_rank', 'lora_alpha', 'lora_dropout']) {
+      assert.match(task, new RegExp(`\\b${field}\\s*:`), `${taskName} 缺少 LoRA 参数：${field}`)
+    }
+  }
+}
+
+const medicalTask = objectFromMarker(mockData, "name: 'medical-cpt-001'")
+assert.match(medicalTask, /\bprocess_id\s*:/, '运行中的医疗训练任务缺少进程 ID')
+assert.match(mockData, /medical-cpt-001_pid28741\.log/, '医疗训练任务缺少 PID 对应的日志文件')
+assert.match(mockData, /export const mockTrainingLogContents/, '训练日志没有按文件提供独立 Mock 内容')
+assert.match(mockData, /Num examples\s*=\s*9,800/, '医疗训练日志缺少样本规模信息')
+assert.match(mockData, /Total optimization steps\s*=\s*921/, '医疗训练日志缺少真实训练步数信息')
+assert.match(mockData, /\\'epoch\\':\s*1\.92/, '医疗训练日志的 Epoch 未与 64% 进度对齐')
+assert.match(mockData, /\\'loss\\':\s*1\.146/, '医疗训练日志缺少当前 Loss 指标')
+assert.match(mockAdapter, /const file = String\(params\.file/, '训练日志接口没有读取请求中的日志文件名')
+assert.match(mockAdapter, /mockTrainingLogContents\[file\]/, '训练日志接口没有按请求文件返回对应内容')
+assert.match(modelSource, /epoch:\s*number\[\]/, '训练指标模型没有保留逐步 Epoch')
+
+console.log('训练日志 Mock 数据回归检查通过')

frontend/scripts/run-regressions.mjs

@@ -0,0 +1,15 @@
+import { readdir } from 'node:fs/promises'
+import { pathToFileURL } from 'node:url'
+import path from 'node:path'
+
+const scriptsDir = path.resolve(process.cwd(), 'scripts')
+const regressionScripts = (await readdir(scriptsDir))
+  .filter((file) => file.startsWith('regression-') && file.endsWith('.mjs'))
+  .sort()
+
+for (const script of regressionScripts) {
+  console.log(`\n▶ ${script}`)
+  await import(pathToFileURL(path.join(scriptsDir, script)).href)
+}
+
+console.log(`\n✅ ${regressionScripts.length} 个前端回归脚本全部通过`)

frontend/src/api/modules/fineTune.ts

@@ -1,5 +1,5 @@
 import { get, post, put, del } from '../request'
-import type { FineTuneTask, TrainingProgress } from '@/types'
+import type { FineTuneStartPayload, FineTuneTask, TrainingProgress } from '@/types'
 
 /** 训练任务列表 */
 export const getFineTuneList = () => get<FineTuneTask[]>('/fine-tune')
@@ -16,7 +16,7 @@ export const createFineTune = (data: Partial<FineTuneTask>) =>
   post<{ id: string | number }>('/fine-tune', data)
 
 /** 启动训练（第二步） */
-export const startFineTune = (data: any) => post('/fine-tune/start', data)
+export const startFineTune = (data: FineTuneStartPayload) => post('/fine-tune/start', data)
 
 /** 更新训练任务 */
 export const updateFineTune = (id: string | number, data: Partial<FineTuneTask>) =>

frontend/src/mock/adapter.ts

@@ -3,7 +3,7 @@
  * 拦截所有 API 请求并返回 mock 数据
  * 通过 URL + method 路由到对应的 mock 响应
  */
-import type { AxiosInstance, AxiosRequestConfig } from 'axios'
+import type { AxiosAdapter, AxiosInstance, AxiosRequestConfig } from 'axios'
 import {
   mockLoginOk,
   mockHealth,
@@ -21,6 +21,7 @@ import {
   mockLogFiles,
   mockTrainingLogFiles,
   mockLogContent,
+  mockTrainingLogContents,
 } from './data'
 import {
   activateDatasetVersion,
@@ -164,7 +165,8 @@ async function handleMock(config: AxiosRequestConfig) {
   }
   m = url.match(/^\/dataset-manage\/([^/]+)$/)
   if (m && method === 'get') {
-    const found = mockDatasets.find((x) => String(x.id) === m[1])
+    const datasetId = m[1]
+    const found = mockDatasets.find((x) => String(x.id) === datasetId)
     return found ? ok(found) : fail('数据集不存在', 404)
   }
   if (m && (method === 'put' || method === 'delete')) {
@@ -261,7 +263,8 @@ async function handleMock(config: AxiosRequestConfig) {
   }
   m = url.match(/^\/fine-tune\/progress\/([^/]+)$/)
   if (m && method === 'get') {
-    const task = mockFineTuneList.find((t) => String(t.id) === m[1])
+    const taskId = m[1]
+    const task = mockFineTuneList.find((t) => String(t.id) === taskId)
     if (!task) return fail('任务不存在', 404)
     if (task.status === 'running') {
       return ok({
@@ -276,7 +279,8 @@ async function handleMock(config: AxiosRequestConfig) {
   }
   m = url.match(/^\/fine-tune\/([^/]+)$/)
   if (m && method === 'get') {
-    const found = mockFineTuneList.find((x) => String(x.id) === m[1])
+    const taskId = m[1]
+    const found = mockFineTuneList.find((x) => String(x.id) === taskId)
     return found ? ok(found) : fail('任务不存在', 404)
   }
   m = url.match(/^\/fine-tune\/stop\/([^/]+)$/)
@@ -296,7 +300,8 @@ async function handleMock(config: AxiosRequestConfig) {
   }
   m = url.match(/^\/model-compare\/([^/]+)$/)
   if (m && method === 'get') {
-    const found = mockCompareList.find((x) => String(x.id) === m[1])
+    const compareId = m[1]
+    const found = mockCompareList.find((x) => String(x.id) === compareId)
     return found ? ok(found) : fail('任务不存在', 404)
   }
   if (m && method === 'delete') return ok({ deleted: m[1] })
@@ -363,7 +368,14 @@ async function handleMock(config: AxiosRequestConfig) {
   if (url === '/log-files' && method === 'get') return ok(mockLogFiles)
   if (url === '/log-content' && method === 'get') return ok(mockLogContent)
   if (url === '/training-log-files' && method === 'get') return ok(mockTrainingLogFiles)
-  if (url === '/training-log-content' && method === 'get') return ok(mockLogContent)
+  if (url === '/training-log-content' && method === 'get') {
+    const file = String(params.file || '')
+    return ok(mockTrainingLogContents[file] || {
+      file,
+      size: '0 KB',
+      content: `[Mock] 未找到训练日志内容：${file}`,
+    })
+  }
 
   // 未匹配的请求 → 兜底返回空成功（避免阻断 UI）
   console.warn('[Mock] 未匹配路由:', method.toUpperCase(), url, params)
@@ -380,12 +392,13 @@ function safeJSON(str: string) {
 
 /** 给 axios instance 安装 mock adapter */
 export function installMockAdapter(instance: AxiosInstance) {
-  instance.defaults.adapter = async (config: AxiosRequestConfig) => {
+  const adapter = async (config: AxiosRequestConfig) => {
     try {
       const response = await handleMock(config)
       return response
-    } catch (e: any) {
+    } catch (error: unknown) {
-      return fail(e.message || 'Mock 错误', 500, config)
+      return fail(error instanceof Error ? error.message : 'Mock 错误', 500, config)
     }
   }
+  instance.defaults.adapter = adapter as AxiosAdapter
 }

frontend/src/mock/data.ts

@@ -58,37 +58,40 @@ export const mockSystemInfo: SystemInfo = {
       id: 0,
       uuid: 'GPU-MOCK-A800-00',
       name: 'NVIDIA A800',
-      status: 'idle',
+      status: 'busy',
-      gpu_percent: 0,
+      gpu_percent: 74,
-      memory_used_gb: 0,
+      memory_used_gb: 41.8,
       memory_total_gb: 80,
-      memory_percent: 0,
+      memory_percent: 52.3,
-      temperature: 32,
+      temperature: 63,
-      power_w: 38,
+      power_w: 286,
       power_limit_w: 400,
-      fan_speed: 0,
+      fan_speed: 51,
       clock_mhz: 1410,
       driver_version: '535.86.10',
-      processes: [],
+      processes: [
+        { pid: 28741, name: 'python', task_name: 'medical-cpt-001 / rank 0', user: 'trainer', memory_used_gb: 39.6 },
+        { pid: 28768, name: 'python', task_name: '训练指标采集', user: 'trainer', memory_used_gb: 2.2 },
+      ],
     },
     {
       id: 1,
       uuid: 'GPU-MOCK-A800-01',
       name: 'NVIDIA A800',
       status: 'busy',
-      gpu_percent: 28,
+      gpu_percent: 71,
-      memory_used_gb: 22.5,
+      memory_used_gb: 42.1,
       memory_total_gb: 80,
-      memory_percent: 28.1,
+      memory_percent: 52.6,
-      temperature: 52,
+      temperature: 62,
-      power_w: 165,
+      power_w: 279,
       power_limit_w: 400,
-      fan_speed: 32,
+      fan_speed: 49,
       clock_mhz: 1410,
       driver_version: '535.86.10',
       processes: [
-        { pid: 18421, name: 'python', task_name: '指令微调任务', user: 'trainer', memory_used_gb: 18.6 },
+        { pid: 28742, name: 'python', task_name: 'medical-cpt-001 / rank 1', user: 'trainer', memory_used_gb: 39.9 },
-        { pid: 18503, name: 'python', task_name: '训练指标采集', user: 'trainer', memory_used_gb: 3.9 },
+        { pid: 28769, name: 'python', task_name: '训练指标采集', user: 'trainer', memory_used_gb: 2.2 },
       ],
     },
     {
@@ -249,7 +252,7 @@ export const mockLocalModels = {
 }
 
 // ============ 数据集 ============
-export const mockDatasets: DatasetItem[] = [
+export const mockDatasets: DatasetItem[] = ([
   {
     id: 1,
     name: '金融问答-训练集',
@@ -273,7 +276,7 @@ export const mockDatasets: DatasetItem[] = [
   { id: 8, name: '通用指令构造集', type: 'train', storage_type: 'local', source: 'task', task_id: 492015, size: '148 MB', count: 12600, description: '由指令微调数据构造任务生成', create_time: '2026-07-09T01:42:00Z' },
   { id: 9, name: '用户反馈脱敏集', type: 'test', storage_type: 'minio', source: 'task', task_id: 731948, size: '72 MB', count: 9340, description: '由敏感信息脱敏任务生成', create_time: '2026-07-09T09:18:00Z' },
   { id: 10, name: '多轮对话增强集', type: 'eval', storage_type: 'local', source: 'task', task_id: 582012, size: '41 MB', count: 2780, description: '由多轮对话拼接任务生成', create_time: '2026-07-10T02:06:00Z' },
-].map((dataset) => ({
+] satisfies DatasetItem[]).map((dataset) => ({
   ...dataset,
   files: dataset.files?.length
     ? dataset.files
@@ -339,12 +342,181 @@ export const mockDatasetPreviews: Record<string, string> = {
 
 // ============ 训练任务 ============
 export const mockFineTuneList: FineTuneTask[] = [
-  { id: 103942, name: 'finance-sft-001', description: '金融领域 SFT 训练', status: 'completed', train_type: 'SFT', train_method: 'lora', template: 'qwen', base_model: 1, train_dataset_id: 1, gpus: [0], progress: 100, train_duration: '2小时18分钟', create_time: '2026-01-15T08:00:00Z' },
+  {
-  { id: 349102, name: 'legal-sft-002', description: '法律文书 SFT', status: 'completed', train_type: 'SFT', train_method: 'lora', template: 'qwen', base_model: 1, train_dataset_id: 2, gpus: [1], progress: 100, train_duration: '1小时46分钟', create_time: '2026-01-18T10:00:00Z' },
+    id: 103942,
-  { id: 849301, name: 'medical-cpt-001', description: '医疗领域继续预训练', status: 'running', train_type: 'CPT', train_method: 'lora', template: 'qwen2_5', base_model: 2, train_dataset_id: 6, gpus: [0, 1], progress: 64, train_duration: '36分钟', create_time: '2026-02-05T09:00:00Z' },
+    name: 'finance-sft-001',
-  { id: 593021, name: 'service-dpo-001', description: '客服对话偏好训练', status: 'pending', train_type: 'DPO', train_method: 'lora', template: 'qwen', base_model: 1, train_dataset_id: 3, gpus: [2], progress: 0, train_duration: '-', create_time: '2026-02-08T14:00:00Z' },
+    description: '金融领域 SFT 训练',
-  { id: 201948, name: 'finance-sft-002', description: '金融领域二轮微调', status: 'failed', train_type: 'SFT', train_method: 'lora', template: 'qwen', base_model: 1, train_dataset_id: 1, gpus: [3], progress: 32, train_duration: '18分钟', create_time: '2026-02-10T11:00:00Z' },
+    status: 'completed',
-  { id: 940212, name: 'general-sft-001', description: '通用能力微调', status: 'completed', train_type: 'SFT', train_method: 'full', template: 'llama3', base_model: 3, train_dataset_id: 3, gpus: [0, 2], progress: 100, train_duration: '3小时05分钟', create_time: '2026-02-12T13:00:00Z' },
+    train_type: 'SFT',
+    train_method: 'lora',
+    template: 'qwen',
+    base_model: 1,
+    train_dataset_id: 1,
+    output_model_name: 'qwen2.5-7b-finance-sft-v1',
+    auto_merge: true,
+    gpus: [0],
+    batch_size: 8,
+    learning_rate: 0.00002,
+    n_epochs: 3,
+    save_steps: 100,
+    lr_scheduler_type: 'cosine',
+    max_length: 2048,
+    warmup_ratio: 0.05,
+    weight_decay: 0.01,
+    lora_rank: 16,
+    lora_alpha: 32,
+    lora_dropout: 0.05,
+    quantization_bit: 0,
+    process_id: 12345,
+    progress: 100,
+    train_duration: '2小时18分钟',
+    create_time: '2026-01-15T08:00:00Z',
+  },
+  {
+    id: 349102,
+    name: 'legal-sft-002',
+    description: '法律文书 SFT',
+    status: 'completed',
+    train_type: 'SFT',
+    train_method: 'lora',
+    template: 'qwen',
+    base_model: 1,
+    train_dataset_id: 2,
+    output_model_name: 'qwen2.5-7b-legal-sft-v2',
+    auto_merge: true,
+    gpus: [1],
+    batch_size: 4,
+    learning_rate: 0.000015,
+    n_epochs: 4,
+    save_steps: 120,
+    lr_scheduler_type: 'linear',
+    max_length: 4096,
+    warmup_ratio: 0.03,
+    weight_decay: 0.01,
+    lora_rank: 32,
+    lora_alpha: 64,
+    lora_dropout: 0.05,
+    quantization_bit: 0,
+    process_id: 12350,
+    progress: 100,
+    train_duration: '1小时46分钟',
+    create_time: '2026-01-18T10:00:00Z',
+  },
+  {
+    id: 849301,
+    name: 'medical-cpt-001',
+    description: '医疗领域继续预训练',
+    status: 'running',
+    train_type: 'CPT',
+    train_method: 'lora',
+    template: 'qwen2_5',
+    base_model: 2,
+    train_dataset_id: 6,
+    output_model_name: 'qwen2.5-14b-medical-cpt-v1',
+    auto_merge: true,
+    gpus: [0, 1],
+    batch_size: 2,
+    learning_rate: 0.0001,
+    n_epochs: 3,
+    save_steps: 200,
+    lr_scheduler_type: 'cosine',
+    max_length: 4096,
+    warmup_ratio: 0.03,
+    weight_decay: 0.01,
+    lora_rank: 16,
+    lora_alpha: 32,
+    lora_dropout: 0.05,
+    quantization_bit: 0,
+    process_id: 28741,
+    progress: 64,
+    train_duration: '36分钟',
+    create_time: '2026-07-13T06:40:00Z',
+  },
+  {
+    id: 593021,
+    name: 'service-dpo-001',
+    description: '客服对话偏好训练',
+    status: 'pending',
+    train_type: 'DPO',
+    train_method: 'lora',
+    template: 'qwen',
+    base_model: 1,
+    train_dataset_id: 3,
+    output_model_name: 'qwen2.5-7b-service-dpo-v1',
+    auto_merge: true,
+    gpus: [2],
+    batch_size: 4,
+    learning_rate: 0.000005,
+    n_epochs: 2,
+    save_steps: 100,
+    lr_scheduler_type: 'cosine',
+    max_length: 2048,
+    warmup_ratio: 0.1,
+    weight_decay: 0,
+    lora_rank: 16,
+    lora_alpha: 32,
+    lora_dropout: 0.1,
+    quantization_bit: 0,
+    progress: 0,
+    train_duration: '等待调度',
+    create_time: '2026-02-08T14:00:00Z',
+  },
+  {
+    id: 201948,
+    name: 'finance-sft-002',
+    description: '金融领域二轮微调',
+    status: 'failed',
+    train_type: 'SFT',
+    train_method: 'lora',
+    template: 'qwen',
+    base_model: 1,
+    train_dataset_id: 1,
+    output_model_name: 'qwen2.5-7b-finance-sft-v2',
+    auto_merge: false,
+    gpus: [3],
+    batch_size: 8,
+    learning_rate: 0.00002,
+    n_epochs: 3,
+    save_steps: 100,
+    lr_scheduler_type: 'cosine',
+    max_length: 2048,
+    warmup_ratio: 0.05,
+    weight_decay: 0.01,
+    lora_rank: 16,
+    lora_alpha: 32,
+    lora_dropout: 0.05,
+    quantization_bit: 0,
+    process_id: 27654,
+    progress: 32,
+    train_duration: '18分钟',
+    create_time: '2026-02-10T11:00:00Z',
+  },
+  {
+    id: 940212,
+    name: 'general-sft-001',
+    description: '通用能力全参数微调',
+    status: 'completed',
+    train_type: 'SFT',
+    train_method: 'full',
+    template: 'llama3',
+    base_model: 3,
+    train_dataset_id: 3,
+    output_model_name: 'llama3-8b-general-sft-v1',
+    auto_merge: false,
+    gpus: [0, 2],
+    batch_size: 2,
+    learning_rate: 0.00001,
+    n_epochs: 2,
+    save_steps: 250,
+    lr_scheduler_type: 'cosine',
+    max_length: 4096,
+    warmup_ratio: 0.03,
+    weight_decay: 0.1,
+    process_id: 26318,
+    progress: 100,
+    train_duration: '3小时05分钟',
+    create_time: '2026-02-12T13:00:00Z',
+  },
 ]
 
 // ============ 模型推理/对比 ============
@@ -549,11 +721,48 @@ export const mockLogFiles: LogFile[] = [
 ]
 
 export const mockTrainingLogFiles: TrainingLogFile[] = [
+  { file: 'medical-cpt-001_pid28741.log', name: 'medical-cpt-001', size: '6.8 MB', pid: 28741, date: '2026-07-13' },
   { file: 'qwen-ft-finance-001_pid12345.log', name: 'finance-sft-001', size: '4.5 MB', pid: 12345, date: '2026-02-15' },
   { file: 'llama3-ft-customer-service_pid12346.log', name: 'service-dpo-001', size: '2.1 MB', pid: 12346, date: '2026-02-18' },
   { file: 'qwen-ft-legal-002_pid12350.log', name: 'legal-sft-002', size: '5.8 MB', pid: 12350, date: '2026-02-20' },
 ]
 
+const medicalTrainingLogLines = [
+  '[2026-07-13 14:40:01] INFO: Launching distributed training with torchrun --nproc_per_node=2',
+  '[2026-07-13 14:40:02] INFO: Process rank: 0, world size: 2, device: cuda:0, distributed training: True',
+  '[2026-07-13 14:40:04] INFO: Loading tokenizer from /data/models/qwen2.5-14b-instruct',
+  '[2026-07-13 14:40:16] INFO: Loading checkpoint shards: 100% | 8/8 | 00:12',
+  '[2026-07-13 14:40:18] INFO: Loading dataset 医疗问答-训练集 (9,800 samples)',
+  '[2026-07-13 14:40:27] INFO: Tokenizing dataset: 100% | 9,800/9,800 | 00:09',
+  '[2026-07-13 14:40:28] INFO: LoRA config: rank=16, alpha=32, dropout=0.05, target_modules=q_proj,k_proj,v_proj,o_proj',
+  '[2026-07-13 14:40:29] INFO: Trainable params: 83,886,080 / 14,787,584,000 (0.5673%)',
+  '[2026-07-13 14:40:30] INFO: ***** Running training *****',
+  '[2026-07-13 14:40:30] INFO:   Num examples = 9,800',
+  '[2026-07-13 14:40:30] INFO:   Num Epochs = 3',
+  '[2026-07-13 14:40:30] INFO:   Instantaneous batch size per device = 2',
+  '[2026-07-13 14:40:30] INFO:   Total train batch size = 32',
+  '[2026-07-13 14:40:30] INFO:   Gradient Accumulation steps = 8',
+  '[2026-07-13 14:40:30] INFO:   Total optimization steps = 921',
+  '[2026-07-13 14:42:18] INFO: step=40 {\'loss\': 2.684, \'grad_norm\': 1.184, \'learning_rate\': 9.82e-05, \'epoch\': 0.13}',
+  '[2026-07-13 14:44:26] INFO: step=80 {\'loss\': 2.312, \'grad_norm\': 1.092, \'learning_rate\': 9.68e-05, \'epoch\': 0.26}',
+  '[2026-07-13 14:46:34] INFO: step=120 {\'loss\': 2.084, \'grad_norm\': 1.037, \'learning_rate\': 9.43e-05, \'epoch\': 0.39}',
+  '[2026-07-13 14:48:42] INFO: step=160 {\'loss\': 1.932, \'grad_norm\': 0.986, \'learning_rate\': 9.08e-05, \'epoch\': 0.52}',
+  '[2026-07-13 14:50:49] INFO: step=200 {\'loss\': 1.801, \'grad_norm\': 0.944, \'learning_rate\': 8.64e-05, \'epoch\': 0.65}',
+  '[2026-07-13 14:50:54] INFO: Saving checkpoint to /data/checkpoints/medical-cpt-001/checkpoint-200',
+  '[2026-07-13 14:52:58] INFO: step=240 {\'loss\': 1.696, \'grad_norm\': 0.913, \'learning_rate\': 8.15e-05, \'epoch\': 0.78}',
+  '[2026-07-13 14:55:06] INFO: step=280 {\'loss\': 1.611, \'grad_norm\': 0.887, \'learning_rate\': 7.61e-05, \'epoch\': 0.91}',
+  '[2026-07-13 14:57:13] INFO: step=320 {\'loss\': 1.532, \'grad_norm\': 0.852, \'learning_rate\': 7.06e-05, \'epoch\': 1.04}',
+  '[2026-07-13 14:59:21] INFO: step=360 {\'loss\': 1.461, \'grad_norm\': 0.829, \'learning_rate\': 6.49e-05, \'epoch\': 1.17}',
+  '[2026-07-13 15:01:29] INFO: step=400 {\'loss\': 1.396, \'grad_norm\': 0.811, \'learning_rate\': 5.91e-05, \'epoch\': 1.30}',
+  '[2026-07-13 15:01:34] INFO: Saving checkpoint to /data/checkpoints/medical-cpt-001/checkpoint-400',
+  '[2026-07-13 15:03:37] INFO: step=440 {\'loss\': 1.337, \'grad_norm\': 0.795, \'learning_rate\': 5.35e-05, \'epoch\': 1.43}',
+  '[2026-07-13 15:05:45] INFO: step=480 {\'loss\': 1.286, \'grad_norm\': 0.776, \'learning_rate\': 4.80e-05, \'epoch\': 1.56}',
+  '[2026-07-13 15:07:53] INFO: step=520 {\'loss\': 1.232, \'grad_norm\': 0.758, \'learning_rate\': 4.28e-05, \'epoch\': 1.69}',
+  '[2026-07-13 15:11:59] INFO: step=560 {\'loss\': 1.184, \'grad_norm\': 0.741, \'learning_rate\': 3.79e-05, \'epoch\': 1.82}',
+  '[2026-07-13 15:16:05] INFO: step=590 {\'loss\': 1.146, \'grad_norm\': 0.728, \'learning_rate\': 3.44e-05, \'epoch\': 1.92}',
+  '[2026-07-13 15:16:06] INFO: Training is running normally, estimated remaining time: 00:20:15',
+]
+
 const fakeLogLines = [
   "[2026-02-15 08:30:12] INFO: Loading model from /data/models/qwen2.5-7b",
   "[2026-02-15 08:30:13] INFO: Loading dataset finance-train-001 (8560 samples)",
@@ -591,3 +800,16 @@ export const mockLogContent: LogContent = {
   size: '2.3 MB',
   content: fakeLogLines.join('\n'),
 }
+
+export const mockTrainingLogContents: Record<string, LogContent> = {
+  'medical-cpt-001_pid28741.log': {
+    file: 'medical-cpt-001_pid28741.log',
+    size: '6.8 MB',
+    content: medicalTrainingLogLines.join('\n'),
+  },
+  'qwen-ft-finance-001_pid12345.log': {
+    file: 'qwen-ft-finance-001_pid12345.log',
+    size: '4.5 MB',
+    content: fakeLogLines.join('\n'),
+  },
+}

frontend/src/types/index.ts

@@ -139,6 +139,37 @@ export interface FineTuneTask {
   create_time?: string
 }
 
+export type FineTuneStartPayload = Omit<
+  FineTuneTask,
+  'id' | 'status' | 'progress' | 'process_id' | 'train_duration' | 'create_time'
+> & {
+  task_id: string | number
+  description: string
+  template: string
+  train_method: TrainMethodType
+  train_dataset_id: string | number
+  auto_merge: boolean
+  output_model_name: string
+  gpus: number[]
+  batch_size: number
+  learning_rate: number
+  n_epochs: number
+  save_steps: number
+  lr_scheduler_type: string
+  max_length: number
+  warmup_ratio: number
+  weight_decay: number
+  lora_alpha: number
+  lora_dropout: number
+  lora_rank: number
+  quantization_bit: number
+  export_quantized: boolean
+  quant_method: string
+  quant_bits: number
+  quant_group_size: number
+  export_format: string
+}
+
 export interface TrainingProgress {
   status?: string
   progress?: number

frontend/src/views/data-process/DataProcessCreateView.vue

@@ -9,17 +9,25 @@ import SourceUploadStep from './create/SourceUploadStep.vue'
 import PreviewCompareStep from './create/PreviewCompareStep.vue'
 import GenerationStep from './create/GenerationStep.vue'
 import ResultEditorStep from './create/ResultEditorStep.vue'
-import { buildPreviewItems, createResults, DEFAULT_SOURCE_TEXT } from './create/previewModel'
+import { buildPreviewItems, DEFAULT_SOURCE_TEXT } from './create/previewModel'
+import {
+  createDefaultStructuredOptions,
+  createDefaultUnstructuredOptions,
+} from './create/dataProcessCreateState'
+import {
+  DATA_PROCESS_DRAFT_STORAGE_KEY,
+  useDataProcessDraft,
+} from './create/useDataProcessDraft'
+import { useDataProcessGeneration } from './create/useDataProcessGeneration'
 import { useModelsStore } from '@/stores/models'
 import type {
   ExternalDataSource,
-  GenerationState,
   PreviewItem,
   ProcessType,
-  ResultItem,
   StepId,
   StructuredProcessOptions,
   UnstructuredProcessOptions,
+  UploadedDataFile,
 } from './create/types'
 
 const router = useRouter()
@@ -28,10 +36,7 @@ const { list: modelList } = storeToRefs(modelsStore)
 const generationModels = computed(() => modelList.value.filter((model) => model.type === 'LLM'))
 const taskSetupRef = ref<InstanceType<typeof TaskSetupStep>>()
 const confirmDialogRef = ref<InstanceType<typeof AppConfirmDialog>>()
-const DRAFT_STORAGE_KEY = 'yg-data-process-create-draft'
-const DRAFT_SCHEMA_VERSION = 5
 const PREVIEW_MODEL_VERSION = 'document-chunk-v2'
-const DEFAULT_GENERATION_PROMPT = '你是一名专业的数据生成助手。请根据输入内容生成准确、完整、可直接用于模型训练的问答数据。仅输出符合目标格式的内容，答案应事实清晰、语言自然，不要添加分析过程、说明或无关内容。'
 
 const WIZARD_STEPS = [
   { id: 'create', title: '创建任务', desc: '填写任务信息与处理配置' },
@@ -40,65 +45,12 @@ const WIZARD_STEPS = [
   { id: 'generate', title: '开始生成', desc: '确认摘要并启动处理' },
   { id: 'results', title: '结果编辑与保存', desc: '检查、修改并保存结果' },
 ] as const satisfies ReadonlyArray<{ id: StepId; title: string; desc: string }>
-const LEGACY_V3_STEP_IDS: ReadonlyArray<StepId> = ['create', 'preview', 'generate', 'results']
-
 const currentStep = ref(0)
 const currentStepId = computed<StepId>(() => WIZARD_STEPS[currentStep.value]?.id ?? 'create')
 const task = reactive({ name: '', description: '' })
 const processType = ref<ProcessType>('unstructured')
-const structuredOptions = ref<StructuredProcessOptions>({
+const structuredOptions = ref<StructuredProcessOptions>(createDefaultStructuredOptions())
-  preprocessOptions: ['clean_invalid', 'detect_structure', 'deduplicate', 'normalize_format'],
+const unstructuredOptions = ref<UnstructuredProcessOptions>(createDefaultUnstructuredOptions())
-  semanticEnrichment: false,
-  qaPairsPerRow: 1,
-  datasetSplit: { train: 80, validation: 10, test: 10 },
-  generationModelId: '',
-  generationPrompt: DEFAULT_GENERATION_PROMPT,
-  temperature: 0.7,
-  maxTokens: 1024,
-  jsonMode: false,
-  qualityFilterEnabled: false,
-  filterLowQuality: true,
-  filterShortContent: true,
-  minOutputLength: 20,
-})
-const unstructuredOptions = ref<UnstructuredProcessOptions>({
-  preprocessOptions: [
-    'clean_invalid_content',
-    'detect_document_structure',
-    'merge_short_content',
-    'filter_low_quality',
-    'deduplicate_content',
-    'preserve_context',
-  ],
-  chunkMethod: 'semantic',
-  chunkSize: 800,
-  chunkOverlap: 100,
-  minChunkSize: 100,
-  customDelimiter: '',
-  preserveTables: true,
-  preserveCodeBlocks: true,
-  preserveLists: true,
-  semanticEnrichment: false,
-  qaPairsPerChunk: 1,
-  datasetSplit: { train: 80, validation: 10, test: 10 },
-  generationModelId: '',
-  generationPrompt: DEFAULT_GENERATION_PROMPT,
-  temperature: 0.7,
-  maxTokens: 1024,
-  jsonMode: false,
-  qualityFilterEnabled: false,
-  filterLowQuality: true,
-  filterShortContent: true,
-  minOutputLength: 20,
-})
-interface UploadedDataFile {
-  uid: number | string
-  name: string
-  size: number
-  count: number
-  content: string
-}
-
 const uploadedFiles = ref<UploadedDataFile[]>([])
 
 const externalSource = reactive<ExternalDataSource>({
@@ -112,35 +64,53 @@ const externalSource = reactive<ExternalDataSource>({
 })
 const externalPulling = ref(false)
 const externalConnected = ref(false)
+let connectionTimer: ReturnType<typeof setTimeout> | null = null
+let pullTimer: ReturnType<typeof setTimeout> | null = null
 
 const fileName = computed(() => uploadedFiles.value.map(f => f.name).join(', '))
-const fileSize = computed(() => uploadedFiles.value.reduce((sum, f) => sum + f.size, 0))
-const fileCount = computed(() => uploadedFiles.value.reduce((sum, f) => sum + f.count, 0))
 const previewSignature = ref('')
 const previewItems = ref<PreviewItem[]>([])
 const selectedPreviewFileId = ref<string | null>(null)
 const selectedPreviewId = ref<string | null>(null)
 const selectedPreviewIdsByFile = ref<Record<string, string>>({})
-const results = ref<ResultItem[]>([])
-const selectedResultId = ref<string | null>(null)
 const dirty = ref(false)
 const restoringDraft = ref(false)
 let allowLeave = false
 
-const generation = reactive<GenerationState>({
+const {
-  status: 'idle',
+  generation,
-  progress: 0,
+  results,
-  message: '确认摘要后即可开始生成，过程中可查看实时进度。',
+  selectedResultId,
+  resetDownstream,
+  restoreResult,
+  startGeneration,
+  stopGeneration,
+  stopGenerationTimer,
+  updateResultField,
+  validateResults,
+} = useDataProcessGeneration({
+  previewItems,
+  processType,
+  structuredOptions,
+  unstructuredOptions,
+  dirty,
 })
 
-let generationTimer: ReturnType<typeof setInterval> | null = null
-
 const modifiedPreviewCount = computed(() => previewItems.value.filter((item) => item.status !== 'original').length)
+const previewItemsByFile = computed(() => {
+  const grouped = new Map<string, PreviewItem[]>()
+  for (const item of previewItems.value) {
+    const items = grouped.get(item.sourceFileId) || []
+    items.push(item)
+    grouped.set(item.sourceFileId, items)
+  }
+  return grouped
+})
 const activePreviewFile = computed(() => uploadedFiles.value.find((file) => String(file.uid) === selectedPreviewFileId.value))
-const activePreviewItems = computed(() => previewItems.value.filter((item) => item.sourceFileId === selectedPreviewFileId.value))
+const activePreviewItems = computed(() => previewItemsByFile.value.get(selectedPreviewFileId.value || '') || [])
 const activeSourceText = computed(() => activePreviewFile.value?.content ?? '')
 const previewFiles = computed(() => uploadedFiles.value.map((file) => {
-  const items = previewItems.value.filter((item) => item.sourceFileId === String(file.uid))
+  const items = previewItemsByFile.value.get(String(file.uid)) || []
   return {
     id: String(file.uid),
     name: file.name,
@@ -169,44 +139,22 @@ const previousStepLabel = computed(() => currentStep.value > 0
   ? WIZARD_STEPS[currentStep.value - 1].title
   : '')
 
-function isStepId(value: unknown): value is StepId {
-  return typeof value === 'string' && WIZARD_STEPS.some((step) => step.id === value)
-}
-
 function goToStep(stepId: StepId) {
   const nextStepIndex = WIZARD_STEPS.findIndex((step) => step.id === stepId)
   if (nextStepIndex >= 0) currentStep.value = nextStepIndex
 }
 
-function draftSnapshot() {
+const { persistDraft, restoreDraft } = useDataProcessDraft({
-  return {
+  currentStepId,
-    schemaVersion: DRAFT_SCHEMA_VERSION,
+  task,
-    currentStep: currentStep.value,
+  processType,
-    currentStepId: currentStepId.value,
+  structuredOptions,
-    task: { ...task },
+  unstructuredOptions,
-    processType: processType.value,
+  externalSource,
-    structuredOptions: {
+  restoringDraft,
-      ...structuredOptions.value,
+  dirty,
-      preprocessOptions: [...structuredOptions.value.preprocessOptions],
+  goToStep,
-      datasetSplit: { ...structuredOptions.value.datasetSplit },
+})
-    },
-    unstructuredOptions: {
-      ...unstructuredOptions.value,
-      preprocessOptions: [...unstructuredOptions.value.preprocessOptions],
-      datasetSplit: { ...unstructuredOptions.value.datasetSplit },
-    },
-    uploadedFiles: uploadedFiles.value,
-    externalSource: { ...externalSource },
-    previewSignature: previewSignature.value,
-    previewItems: previewItems.value,
-    selectedPreviewFileId: selectedPreviewFileId.value,
-    selectedPreviewId: selectedPreviewId.value,
-    selectedPreviewIdsByFile: selectedPreviewIdsByFile.value,
-    results: results.value,
-    selectedResultId: selectedResultId.value,
-    generation: { ...generation },
-  }
-}
 
 function previewAffectingOptions() {
   if (processType.value === 'structured') {
@@ -291,191 +239,11 @@ const generationOptionsSignature = computed(() => JSON.stringify(generationAffec
 
 function buildPreviewSignature() {
   const filesSignature = uploadedFiles.value
-    .map((file) => `${file.uid}:${file.content}`)
+    .map((file) => `${file.uid}:${file.name}:${file.size}:${file.count}`)
     .join('|')
   return `${PREVIEW_MODEL_VERSION}:${processType.value}:${JSON.stringify(previewAffectingOptions())}:${filesSignature}`
 }
 
-function persistDraft() {
-  if (restoringDraft.value) return
-  try {
-    localStorage.setItem(DRAFT_STORAGE_KEY, JSON.stringify(draftSnapshot()))
-  } catch {
-    ElMessage.warning('草稿保存失败，请检查浏览器存储空间')
-  }
-}
-
-type DraftSnapshot = Partial<ReturnType<typeof draftSnapshot>> & {
-  fileName?: string
-  fileSize?: number
-  fileCount?: number
-  sourceText?: string
-}
-
-function restoreDraft() {
-  try {
-    const raw = localStorage.getItem(DRAFT_STORAGE_KEY)
-    if (!raw) return
-    const snapshot = JSON.parse(raw) as DraftSnapshot
-    if (!snapshot?.uploadedFiles && (!snapshot?.fileName || !snapshot?.sourceText)) return
-    const snapshotSchemaVersion = Number(snapshot.schemaVersion) || 0
-    const requiresPreviewMigration = snapshotSchemaVersion < 3
-    const legacyStepId = snapshotSchemaVersion === 3
-      ? LEGACY_V3_STEP_IDS[Math.min(Math.max(Number(snapshot.currentStep) || 0, 0), LEGACY_V3_STEP_IDS.length - 1)]
-      : undefined
-    const indexedStepId = WIZARD_STEPS[Math.min(
-      Math.max(Number(snapshot.currentStep) || 0, 0),
-      WIZARD_STEPS.length - 1,
-    )]?.id
-    const restoredStepId = isStepId(snapshot.currentStepId)
-      ? snapshot.currentStepId
-      : legacyStepId || indexedStepId || 'create'
-
-    restoringDraft.value = true
-    goToStep(requiresPreviewMigration ? 'create' : restoredStepId)
-    task.name = snapshot.task?.name || ''
-    task.description = snapshot.task?.description || ''
-    processType.value = snapshot.processType === 'structured' || snapshot.processType === 'external'
-      ? snapshot.processType
-      : 'unstructured'
-    if (snapshot.structuredOptions) {
-      structuredOptions.value = {
-        ...structuredOptions.value,
-        ...snapshot.structuredOptions,
-        preprocessOptions: Array.isArray(snapshot.structuredOptions.preprocessOptions)
-          ? snapshot.structuredOptions.preprocessOptions
-          : structuredOptions.value.preprocessOptions,
-        datasetSplit: {
-          ...structuredOptions.value.datasetSplit,
-          ...snapshot.structuredOptions.datasetSplit,
-        },
-      }
-    }
-    if (snapshot.unstructuredOptions) {
-      const restoredUnstructuredOptions = snapshot.unstructuredOptions
-      const restoredDatasetSplit = restoredUnstructuredOptions.datasetSplit
-      unstructuredOptions.value = {
-        ...unstructuredOptions.value,
-        preprocessOptions: Array.isArray(restoredUnstructuredOptions.preprocessOptions)
-          ? restoredUnstructuredOptions.preprocessOptions
-          : unstructuredOptions.value.preprocessOptions,
-        chunkMethod: restoredUnstructuredOptions.chunkMethod || unstructuredOptions.value.chunkMethod,
-        chunkSize: Number.isFinite(restoredUnstructuredOptions.chunkSize)
-          ? restoredUnstructuredOptions.chunkSize
-          : unstructuredOptions.value.chunkSize,
-        chunkOverlap: Number.isFinite(restoredUnstructuredOptions.chunkOverlap)
-          ? restoredUnstructuredOptions.chunkOverlap
-          : unstructuredOptions.value.chunkOverlap,
-        minChunkSize: Number.isFinite(restoredUnstructuredOptions.minChunkSize)
-          ? restoredUnstructuredOptions.minChunkSize
-          : unstructuredOptions.value.minChunkSize,
-        customDelimiter: typeof restoredUnstructuredOptions.customDelimiter === 'string'
-          ? restoredUnstructuredOptions.customDelimiter
-          : unstructuredOptions.value.customDelimiter,
-        preserveTables: typeof restoredUnstructuredOptions.preserveTables === 'boolean'
-          ? restoredUnstructuredOptions.preserveTables
-          : unstructuredOptions.value.preserveTables,
-        preserveCodeBlocks: typeof restoredUnstructuredOptions.preserveCodeBlocks === 'boolean'
-          ? restoredUnstructuredOptions.preserveCodeBlocks
-          : unstructuredOptions.value.preserveCodeBlocks,
-        preserveLists: typeof restoredUnstructuredOptions.preserveLists === 'boolean'
-          ? restoredUnstructuredOptions.preserveLists
-          : unstructuredOptions.value.preserveLists,
-        semanticEnrichment: typeof restoredUnstructuredOptions.semanticEnrichment === 'boolean'
-          ? restoredUnstructuredOptions.semanticEnrichment
-          : unstructuredOptions.value.semanticEnrichment,
-        qaPairsPerChunk: Number.isFinite(restoredUnstructuredOptions.qaPairsPerChunk)
-          ? restoredUnstructuredOptions.qaPairsPerChunk
-          : unstructuredOptions.value.qaPairsPerChunk,
-        generationModelId: restoredUnstructuredOptions.generationModelId ?? unstructuredOptions.value.generationModelId,
-        generationPrompt: typeof restoredUnstructuredOptions.generationPrompt === 'string'
-          ? restoredUnstructuredOptions.generationPrompt
-          : unstructuredOptions.value.generationPrompt,
-        temperature: Number.isFinite(restoredUnstructuredOptions.temperature)
-          ? restoredUnstructuredOptions.temperature
-          : unstructuredOptions.value.temperature,
-        maxTokens: Number.isFinite(restoredUnstructuredOptions.maxTokens)
-          ? restoredUnstructuredOptions.maxTokens
-          : unstructuredOptions.value.maxTokens,
-        jsonMode: typeof restoredUnstructuredOptions.jsonMode === 'boolean'
-          ? restoredUnstructuredOptions.jsonMode
-          : unstructuredOptions.value.jsonMode,
-        qualityFilterEnabled: typeof restoredUnstructuredOptions.qualityFilterEnabled === 'boolean'
-          ? restoredUnstructuredOptions.qualityFilterEnabled
-          : unstructuredOptions.value.qualityFilterEnabled,
-        filterLowQuality: typeof restoredUnstructuredOptions.filterLowQuality === 'boolean'
-          ? restoredUnstructuredOptions.filterLowQuality
-          : unstructuredOptions.value.filterLowQuality,
-        filterShortContent: typeof restoredUnstructuredOptions.filterShortContent === 'boolean'
-          ? restoredUnstructuredOptions.filterShortContent
-          : unstructuredOptions.value.filterShortContent,
-        minOutputLength: Number.isFinite(restoredUnstructuredOptions.minOutputLength)
-          ? restoredUnstructuredOptions.minOutputLength
-          : unstructuredOptions.value.minOutputLength,
-        datasetSplit: {
-          train: Number.isFinite(restoredDatasetSplit?.train)
-            ? restoredDatasetSplit.train
-            : unstructuredOptions.value.datasetSplit.train,
-          validation: Number.isFinite(restoredDatasetSplit?.validation)
-            ? restoredDatasetSplit.validation
-            : unstructuredOptions.value.datasetSplit.validation,
-          test: Number.isFinite(restoredDatasetSplit?.test)
-            ? restoredDatasetSplit.test
-            : unstructuredOptions.value.datasetSplit.test,
-        },
-      }
-    }
-    
-    if (snapshot.uploadedFiles) {
-      uploadedFiles.value = Array.isArray(snapshot.uploadedFiles) ? snapshot.uploadedFiles : []
-    } else if (snapshot.fileName && snapshot.sourceText) {
-      // Migrate old draft
-      uploadedFiles.value = [{
-        uid: 'migrated-draft',
-        name: snapshot.fileName,
-        size: Number(snapshot.fileSize) || 0,
-        count: Number(snapshot.fileCount) || 0,
-        content: snapshot.sourceText
-      }]
-    }
-
-    previewSignature.value = requiresPreviewMigration ? '' : snapshot.previewSignature || ''
-    if (snapshot.externalSource) {
-      Object.assign(externalSource, snapshot.externalSource)
-    }
-    const defaultSourceFileId = String(uploadedFiles.value[0]?.uid ?? '')
-    previewItems.value = !requiresPreviewMigration && Array.isArray(snapshot.previewItems)
-      ? snapshot.previewItems.map((item) => ({ ...item, sourceFileId: item.sourceFileId ?? defaultSourceFileId }))
-      : []
-    selectedPreviewFileId.value = snapshot.selectedPreviewFileId || defaultSourceFileId || null
-    selectedPreviewIdsByFile.value = requiresPreviewMigration ? {} : snapshot.selectedPreviewIdsByFile || {}
-    selectedPreviewId.value = requiresPreviewMigration
-      ? null
-      : snapshot.selectedPreviewId
-        || selectedPreviewIdsByFile.value[selectedPreviewFileId.value ?? '']
-        || activePreviewItems.value[0]?.id
-        || null
-    results.value = !requiresPreviewMigration && Array.isArray(snapshot.results) ? snapshot.results : []
-    selectedResultId.value = requiresPreviewMigration ? null : snapshot.selectedResultId || results.value[0]?.id || null
-    if (!requiresPreviewMigration) {
-      const restoredGeneration = snapshot.generation || {}
-      Object.assign(generation, restoredGeneration)
-      if (generation.status === 'running') {
-        generation.status = 'idle'
-        generation.progress = 0
-        generation.message = '草稿已恢复，请重新开始生成。'
-      }
-    }
-    dirty.value = false
-    nextTick(() => { restoringDraft.value = false })
-    ElMessage.info(requiresPreviewMigration
-      ? '已恢复任务信息，切分规则已更新，请重新生成预览'
-      : '已恢复上次保存的草稿')
-  } catch {
-    localStorage.removeItem(DRAFT_STORAGE_KEY)
-  }
-}
-
 watch(
   [() => task.name, () => task.description, processType, structuredOptions, unstructuredOptions, externalSource],
   () => {
@@ -496,9 +264,7 @@ watch(generationOptionsSignature, (currentSignature, previousSignature) => {
 })
 
 watch(
-  [currentStep, task, processType, structuredOptions, unstructuredOptions, uploadedFiles, externalSource, previewSignature,
+  [task, processType, structuredOptions, unstructuredOptions, externalSource],
-    previewItems, selectedPreviewFileId, selectedPreviewId, selectedPreviewIdsByFile,
-    results, selectedResultId, generation],
   persistDraft,
   { deep: true },
 )
@@ -566,8 +332,10 @@ function handleTestConnection() {
     ElMessage.warning('请先填写数据源地址')
     return
   }
+  if (connectionTimer) clearTimeout(connectionTimer)
   externalPulling.value = true
-  setTimeout(() => {
+  connectionTimer = setTimeout(() => {
+    connectionTimer = null
     externalPulling.value = false
     externalConnected.value = true
     ElMessage.success('数据源连接测试成功')
@@ -579,8 +347,10 @@ function handlePullData() {
     ElMessage.warning('请先填写数据源地址')
     return
   }
+  if (pullTimer) clearTimeout(pullTimer)
   externalPulling.value = true
-  setTimeout(() => {
+  pullTimer = setTimeout(() => {
+    pullTimer = null
     externalPulling.value = false
     externalConnected.value = true
     const typeName = externalSource.type.toUpperCase()
@@ -620,15 +390,6 @@ function resetSourceDataForProcessTypeChange() {
   resetDownstream()
 }
 
-function resetDownstream() {
-  stopGenerationTimer()
-  generation.status = 'idle'
-  generation.progress = 0
-  generation.message = '确认摘要后即可开始生成，过程中可查看实时进度。'
-  results.value = []
-  selectedResultId.value = null
-}
-
 async function nextFromCreate() {
   const valid = await taskSetupRef.value?.validate()
   if (!valid) return
@@ -738,81 +499,6 @@ async function removePreviewItem(id: string) {
   dirty.value = true
 }
 
-function stopGenerationTimer() {
-  if (generationTimer) clearInterval(generationTimer)
-  generationTimer = null
-}
-
-function startGeneration() {
-  stopGenerationTimer()
-  generation.status = 'running'
-  generation.progress = 0
-  generation.message = '正在应用预览修改并生成标准化结果，请稍候。'
-
-  generationTimer = setInterval(() => {
-    generation.progress = Math.min(100, generation.progress + 8)
-    if (generation.progress < 100) return
-
-    stopGenerationTimer()
-    generation.status = 'success'
-    results.value = createResults(
-      previewItems.value,
-      processType.value === 'structured'
-        ? structuredOptions.value
-        : processType.value === 'unstructured'
-          ? unstructuredOptions.value
-          : undefined,
-    )
-    generation.message = `已完成 ${results.value.length.toLocaleString()} 条数据处理，可进入结果页检查。`
-    selectedResultId.value = results.value[0]?.id ?? null
-    dirty.value = true
-    ElMessage.success('数据处理完成')
-  }, 180)
-}
-
-function stopGeneration() {
-  stopGenerationTimer()
-  generation.status = 'failed'
-  generation.message = '任务已停止，预览修改仍然保留，可以重新生成。'
-}
-
-function updateResultField(id: string, field: 'instruction' | 'input' | 'output', value: string) {
-  const item = results.value.find((entry) => entry.id === id)
-  if (!item) return
-  item[field] = value
-  const valid = item.instruction.trim() && item.output.trim()
-  item.error = valid ? undefined : 'Instruction 和 Output 不能为空'
-  const changed = item.instruction !== item.originalInstruction
-    || item.input !== item.originalInput
-    || item.output !== item.originalOutput
-  item.status = item.error ? 'invalid' : changed ? 'modified' : 'valid'
-  dirty.value = true
-}
-
-function restoreResult(id: string) {
-  const item = results.value.find((entry) => entry.id === id)
-  if (!item) return
-  item.instruction = item.originalInstruction
-  item.input = item.originalInput
-  item.output = item.originalOutput
-  item.error = undefined
-  item.status = 'valid'
-  dirty.value = true
-}
-
-function validateResults() {
-  let firstInvalidId: string | null = null
-  for (const item of results.value) {
-    if (!item.instruction.trim() || !item.output.trim()) {
-      item.error = 'Instruction 和 Output 不能为空'
-      item.status = 'invalid'
-      firstInvalidId ??= item.id
-    }
-  }
-  if (firstInvalidId) selectedResultId.value = firstInvalidId
-  return firstInvalidId == null
-}
-
 async function handlePrimaryAction() {
   if (currentStepId.value === 'create') {
     await nextFromCreate()
@@ -856,7 +542,7 @@ async function saveTask() {
     return
   }
   dirty.value = false
-  localStorage.removeItem(DRAFT_STORAGE_KEY)
+  localStorage.removeItem(DATA_PROCESS_DRAFT_STORAGE_KEY)
   allowLeave = true
   ElMessage.success('数据处理任务已保存')
   await router.push('/data-process')
@@ -893,7 +579,11 @@ onBeforeRouteLeave(async () => {
   return Boolean(confirmed)
 })
 
-onBeforeUnmount(stopGenerationTimer)
+onBeforeUnmount(() => {
+  stopGenerationTimer()
+  if (connectionTimer) clearTimeout(connectionTimer)
+  if (pullTimer) clearTimeout(pullTimer)
+})
 onMounted(() => {
   restoreDraft()
   modelsStore.load()
@@ -1023,198 +713,4 @@ onMounted(() => {
   <AppConfirmDialog ref="confirmDialogRef" />
 </template>
 
-<style scoped lang="scss">
+<style scoped lang="scss" src="./create/data-process-create.scss"></style>
-.create-wizard-layout {
-  display: flex;
-  flex-direction: column;
-  height: 100%;
-  border: 1px solid #eef0f5;
-  border-radius: 8px;
-  background: #fff;
-}
-
-.wizard-main {
-  flex: 1;
-  overflow-y: auto;
-  padding: 32px;
-  
-  /* 自定义滚动条 */
-  &::-webkit-scrollbar {
-    width: 6px;
-  }
-  &::-webkit-scrollbar-thumb {
-    background: #cbd5e1;
-    border-radius: 3px;
-  }
-}
-
-.wizard-main-inner {
-  min-height: 100%;
-  box-sizing: border-box;
-}
-
-.wizard-steps-container {
-  margin-bottom: 32px;
-  padding-bottom: 24px;
-  border-bottom: 1px dashed #e2e8f0;
-}
-
-.custom-wizard-steps {
-  display: flex;
-  align-items: center;
-  justify-content: space-between;
-  max-width: 860px;
-  margin: 0 auto;
-  padding: 0 20px;
-}
-
-.step-item {
-  display: flex;
-  align-items: center;
-  flex: 1;
-
-  &:first-child {
-    flex: 0;
-  }
-}
-
-.step-connector {
-  flex: 1;
-  height: 2px;
-  background-color: #e2e8f0;
-  margin: 0 16px;
-  transition: background-color 0.3s;
-}
-
-.step-item.is-completed .step-connector,
-.step-item.is-active .step-connector {
-  background-color: #5146e5;
-}
-
-.step-node {
-  display: flex;
-  align-items: center;
-  gap: 10px;
-}
-
-.step-icon {
-  display: flex;
-  align-items: center;
-  justify-content: center;
-  width: 28px;
-  height: 28px;
-  border-radius: 50%;
-  border: 2px solid #cbd5e1;
-  background-color: #fff;
-  color: #64748b;
-  font-size: 13px;
-  font-weight: 650;
-  transition: all 0.3s;
-}
-
-.step-title {
-  font-size: 14px;
-  font-weight: 600;
-  color: #64748b;
-  white-space: nowrap;
-}
-
-/* State: Active */
-.step-item.is-active {
-  .step-icon {
-    border-color: #5146e5;
-    background-color: #eef2ff;
-    color: #5146e5;
-  }
-  .step-title {
-    color: #1e293b;
-  }
-}
-
-/* State: Completed */
-.step-item.is-completed {
-  .step-icon {
-    background-color: #5146e5;
-    border-color: #5146e5;
-    color: #fff;
-  }
-  .step-title {
-    color: #1e293b;
-  }
-}
-
-.wizard-content {
-  min-height: 400px;
-}
-
-.wizard-footer {
-  flex-shrink: 0;
-  height: 64px;
-  background: #ffffff;
-  border-top: 1px solid #e2e8f0;
-  display: grid;
-  grid-template-columns: 1fr auto 1fr;
-  align-items: center;
-  padding: 0 32px;
-  box-shadow: 0 -4px 6px -1px rgba(0, 0, 0, 0.02);
-
-  .footer-left {
-    justify-self: start;
-  }
-  
-  .footer-center {
-    justify-self: center;
-  }
-
-  .footer-right {
-    justify-self: end;
-  }
-}
-
-.wizard-primary-action {
-  min-width: 160px;
-}
-
-@media (max-width: 760px) {
-  .wizard-main {
-    padding: 24px 20px;
-  }
-
-  .custom-wizard-steps {
-    padding: 0;
-  }
-
-  .step-connector {
-    margin: 0 8px;
-  }
-
-  .step-node {
-    gap: 6px;
-  }
-
-  .step-title {
-    max-width: 72px;
-    font-size: 12px;
-    line-height: 1.35;
-    white-space: normal;
-  }
-
-  .wizard-footer {
-    padding: 0 20px;
-  }
-}
-
-@media (max-width: 600px) {
-  .step-title {
-    display: none;
-  }
-
-  .step-connector {
-    margin: 0 6px;
-  }
-
-  .wizard-primary-action {
-    min-width: 0;
-  }
-}
-</style>

frontend/src/views/data-process/create/DatasetSplitEditor.vue

@@ -0,0 +1,185 @@
+<script setup lang="ts">
+import { computed } from 'vue'
+import type { DatasetSplitOptions } from './types'
+
+const props = withDefaults(defineProps<{
+  modelValue: DatasetSplitOptions
+  ariaLabelPrefix?: string
+}>(), {
+  ariaLabelPrefix: '',
+})
+
+const emit = defineEmits<{
+  'update:modelValue': [value: DatasetSplitOptions]
+}>()
+
+const splitTotal = computed(() => {
+  const { train, validation, test } = props.modelValue
+  return train + validation + test
+})
+
+function updateField(field: keyof DatasetSplitOptions, value: number | undefined) {
+  emit('update:modelValue', {
+    ...props.modelValue,
+    [field]: Math.min(100, Math.max(0, Number(value) || 0)),
+  })
+}
+
+function ariaLabel(label: string) {
+  return `${props.ariaLabelPrefix}${label}比例`
+}
+</script>
+
+<template>
+  <div class="generation-option-row dataset-split-row">
+    <div class="generation-option-copy">
+      <strong>数据集划分</strong>
+      <small>按比例将生成后的问答对随机划分为训练集、验证集和测试集</small>
+    </div>
+    <div class="dataset-split-config">
+      <div class="dataset-split-grid">
+        <label class="dataset-split-field">
+          <span>训练集</span>
+          <span class="dataset-split-input">
+            <el-input-number
+              :model-value="modelValue.train"
+              :min="0"
+              :max="100"
+              :step="1"
+              :precision="0"
+              controls-position="right"
+              :aria-label="ariaLabel('训练集')"
+              @update:model-value="updateField('train', $event)"
+            />
+            <span>%</span>
+          </span>
+        </label>
+        <label class="dataset-split-field">
+          <span>验证集</span>
+          <span class="dataset-split-input">
+            <el-input-number
+              :model-value="modelValue.validation"
+              :min="0"
+              :max="100"
+              :step="1"
+              :precision="0"
+              controls-position="right"
+              :aria-label="ariaLabel('验证集')"
+              @update:model-value="updateField('validation', $event)"
+            />
+            <span>%</span>
+          </span>
+        </label>
+        <label class="dataset-split-field">
+          <span>测试集</span>
+          <span class="dataset-split-input">
+            <el-input-number
+              :model-value="modelValue.test"
+              :min="0"
+              :max="100"
+              :step="1"
+              :precision="0"
+              controls-position="right"
+              :aria-label="ariaLabel('测试集')"
+              @update:model-value="updateField('test', $event)"
+            />
+            <span>%</span>
+          </span>
+        </label>
+      </div>
+      <div class="dataset-split-summary" :class="{ 'is-invalid': splitTotal !== 100 }">
+        <span>总计 {{ splitTotal }}%</span>
+        <span v-if="splitTotal !== 100" role="alert">
+          训练集、验证集和测试集比例总和必须为 100%
+        </span>
+      </div>
+    </div>
+  </div>
+</template>
+
+<style scoped lang="scss">
+.generation-option-row {
+  display: flex;
+  min-height: 64px;
+  align-items: center;
+  justify-content: space-between;
+  gap: 24px;
+  padding: 12px 14px;
+  box-sizing: border-box;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+}
+
+.generation-option-copy {
+  display: grid;
+  gap: 4px;
+
+  strong {
+    color: #344054;
+    font-size: 13px;
+    font-weight: 600;
+  }
+
+  small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
+  }
+}
+
+.dataset-split-row {
+  align-items: stretch;
+  flex-direction: column;
+}
+
+.dataset-split-config {
+  display: grid;
+  gap: 10px;
+}
+
+.dataset-split-grid {
+  display: grid;
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+  gap: 16px;
+}
+
+.dataset-split-field {
+  display: grid;
+  gap: 7px;
+  color: #5f6878;
+  font-size: 12px;
+}
+
+.dataset-split-input {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+
+  :deep(.el-input-number) {
+    width: 100%;
+  }
+}
+
+.dataset-split-summary {
+  display: flex;
+  justify-content: space-between;
+  gap: 12px;
+  color: #2ca66a;
+  font-size: 12px;
+
+  &.is-invalid {
+    color: #dc2626;
+  }
+}
+
+@media (max-width: 900px) {
+  .dataset-split-grid {
+    grid-template-columns: minmax(0, 1fr);
+  }
+
+  .dataset-split-summary {
+    align-items: flex-start;
+    flex-direction: column;
+  }
+}
+</style>

frontend/src/views/data-process/create/GenerationOptionsPanel.vue

@@ -95,7 +95,7 @@ function sectionValidationMessage() {
         @update:model-value="updateField('generationPrompt', $event)"
       />
       <div class="prompt-variables-hint">
-        提示：可在文本中通过 <code>{{ content }}</code> 引用当前正在处理的数据内容。
+        提示：可在文本中通过 <code v-text="'{{ content }}'" /> 引用当前正在处理的数据内容。
       </div>
     </div>
 

frontend/src/views/data-process/create/SourceUploadStep.vue

@@ -1,18 +1,11 @@
 <script setup lang="ts">
 import { computed, ref, watch } from 'vue'
 import type { UploadFile } from 'element-plus'
-import type { ExternalDataSource, ProcessType } from './types'
+import type { ExternalDataSource, ProcessType, UploadedDataFile } from './types'
-
-interface UploadedSourceFile {
-  uid: string | number
-  name: string
-  size: number
-  count: number
-}
 
 const props = defineProps<{
   processType: ProcessType
-  uploadedFiles: UploadedSourceFile[]
+  uploadedFiles: UploadedDataFile[]
   externalSource: ExternalDataSource
   externalPulling: boolean
   externalConnected: boolean

frontend/src/views/data-process/create/StructuredOptionsPanel.vue

@@ -0,0 +1,270 @@
+<script setup lang="ts">
+import type { ModelItem } from '@/types'
+import type {
+  GenerationControlOptions,
+  PreprocessOption,
+  StructuredProcessOptions,
+} from './types'
+import DatasetSplitEditor from './DatasetSplitEditor.vue'
+import GenerationOptionsPanel from './GenerationOptionsPanel.vue'
+
+const props = defineProps<{
+  options: StructuredProcessOptions
+  generationModels: ModelItem[]
+  validationAttempted: boolean
+  generationValidationMessage: string
+}>()
+
+const emit = defineEmits<{
+  'update:options': [value: StructuredProcessOptions]
+}>()
+
+const PREPROCESS_OPTIONS: Array<{
+  value: PreprocessOption
+  label: string
+  description: string
+}> = [
+  { value: 'clean_invalid', label: '清理无效数据', description: '处理空行、空列和残缺行' },
+  { value: 'detect_structure', label: '识别表格结构', description: '识别表头、多级表头和合并单元格' },
+  { value: 'deduplicate', label: '重复数据去重', description: '删除完全重复或关键字段重复的数据' },
+  { value: 'normalize_format', label: '数据格式标准化', description: '统一日期、数字、单位和枚举值格式' },
+  { value: 'filter_anomaly', label: '异常数据过滤', description: '过滤乱码、无效内容和异常记录' },
+  { value: 'desensitize', label: '敏感信息脱敏', description: '处理姓名、手机号、邮箱等敏感信息' },
+]
+
+function updateField<K extends keyof StructuredProcessOptions>(
+  field: K,
+  value: StructuredProcessOptions[K],
+) {
+  emit('update:options', { ...props.options, [field]: value })
+}
+
+function updateGenerationOptions(value: GenerationControlOptions) {
+  emit('update:options', { ...props.options, ...value })
+}
+
+function updatePreprocessOptions(value: Array<string | number | boolean>) {
+  const allowedValues = new Set(PREPROCESS_OPTIONS.map((option) => option.value))
+  const preprocessOptions = value.filter(
+    (option): option is PreprocessOption => typeof option === 'string' && allowedValues.has(option as PreprocessOption),
+  )
+  updateField('preprocessOptions', preprocessOptions)
+}
+</script>
+
+<template>
+  <div class="form-section structured-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>预处理选项</h3>
+        <p>选择在生成问答对之前需要执行的数据处理方式</p>
+      </div>
+    </div>
+    <el-checkbox-group
+      :model-value="options.preprocessOptions"
+      class="preprocess-option-grid"
+      @update:model-value="updatePreprocessOptions"
+    >
+      <el-checkbox
+        v-for="option in PREPROCESS_OPTIONS"
+        :key="option.value"
+        :value="option.value"
+        class="preprocess-option"
+      >
+        <span class="preprocess-option-copy">
+          <strong>{{ option.label }}</strong>
+          <small>{{ option.description }}</small>
+        </span>
+      </el-checkbox>
+    </el-checkbox-group>
+  </div>
+
+  <div class="form-section generation-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>生成选项</h3>
+        <p>配置每条结构化记录生成问答对的方式和数量</p>
+      </div>
+    </div>
+    <div class="generation-option-list">
+      <GenerationOptionsPanel
+        :options="options"
+        :models="generationModels"
+        section="quality"
+        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        @update:options="updateGenerationOptions"
+      />
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>语义丰富表达</strong>
+          <small>使用大模型将问答表述得更自然、柔和</small>
+        </div>
+        <el-switch
+          :model-value="options.semanticEnrichment"
+          inline-prompt
+          active-text="开"
+          inactive-text="关"
+          @update:model-value="updateField('semanticEnrichment', Boolean($event))"
+        />
+      </div>
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>每行生成数量</strong>
+          <small>每行结构化数据生成的问答对数量</small>
+        </div>
+        <el-input-number
+          :model-value="options.qaPairsPerRow"
+          :min="1"
+          :max="5"
+          :step="1"
+          controls-position="right"
+          @update:model-value="updateField('qaPairsPerRow', Number($event) || 1)"
+        />
+      </div>
+      <DatasetSplitEditor
+        :model-value="options.datasetSplit"
+        @update:model-value="updateField('datasetSplit', $event)"
+      />
+    </div>
+  </div>
+
+  <div class="form-section model-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>大模型</h3>
+        <p>选择数据生成模型，并设置模型输出内容的要求</p>
+      </div>
+    </div>
+    <div class="generation-option-list">
+      <GenerationOptionsPanel
+        :options="options"
+        :models="generationModels"
+        section="model"
+        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        @update:options="updateGenerationOptions"
+      />
+    </div>
+  </div>
+</template>
+
+<style scoped lang="scss">
+.form-section {
+  padding: 0 0 26px;
+  margin-bottom: 26px;
+  border-bottom: 1px solid #edf0f5;
+
+  &:last-child {
+    margin-bottom: 0;
+    border-bottom: 0;
+  }
+
+  h3 {
+    margin: 0 0 16px;
+    color: #2f3747;
+    font-size: 15px;
+    font-weight: 650;
+  }
+}
+
+.section-title-row {
+  display: flex;
+  align-items: flex-start;
+  justify-content: space-between;
+  gap: 16px;
+
+  h3 {
+    margin-bottom: 5px;
+  }
+
+  p {
+    margin: 0;
+    color: #8a93a3;
+    font-size: 12px;
+  }
+}
+
+.preprocess-option-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 12px;
+  margin-top: 16px;
+}
+
+.preprocess-option {
+  width: 100%;
+  min-height: 64px;
+  margin-right: 0;
+  padding: 12px 14px;
+  box-sizing: border-box;
+  align-items: flex-start;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+  transition: border-color 0.18s ease, background-color 0.18s ease;
+
+  &:hover {
+    border-color: #b7b2f7;
+  }
+
+  &.is-checked {
+    background: #fafaff;
+    border-color: #8b82f4;
+  }
+
+  :deep(.el-checkbox__input) {
+    margin-top: 3px;
+  }
+
+  :deep(.el-checkbox__label) {
+    min-width: 0;
+    padding-left: 10px;
+    white-space: normal;
+  }
+}
+
+.preprocess-option-copy,
+.generation-option-copy {
+  display: grid;
+  gap: 4px;
+
+  strong {
+    color: #344054;
+    font-size: 13px;
+    font-weight: 600;
+  }
+
+  small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
+  }
+}
+
+.generation-option-list {
+  display: grid;
+  gap: 12px;
+  margin-top: 16px;
+}
+
+.generation-option-row {
+  display: flex;
+  min-height: 64px;
+  align-items: center;
+  justify-content: space-between;
+  gap: 24px;
+  padding: 12px 14px;
+  box-sizing: border-box;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+}
+
+@media (max-width: 900px) {
+  .preprocess-option-grid {
+    grid-template-columns: minmax(0, 1fr);
+  }
+
+  .generation-option-row {
+    align-items: flex-start;
+    flex-direction: column;
+  }
+}
+</style>

frontend/src/views/data-process/create/UnstructuredOptionsPanel.vue

@@ -0,0 +1,604 @@
+<script setup lang="ts">
+import { computed, ref, watch } from 'vue'
+import type { ModelItem } from '@/types'
+import type {
+  ChunkMethod,
+  GenerationControlOptions,
+  UnstructuredPreprocessOption,
+  UnstructuredProcessOptions,
+} from './types'
+import DatasetSplitEditor from './DatasetSplitEditor.vue'
+import GenerationOptionsPanel from './GenerationOptionsPanel.vue'
+
+const props = defineProps<{
+  options: UnstructuredProcessOptions
+  generationModels: ModelItem[]
+  validationAttempted: boolean
+  generationValidationMessage: string
+  chunkValidationMessage: string
+}>()
+
+const emit = defineEmits<{
+  'update:options': [value: UnstructuredProcessOptions]
+}>()
+
+const SMART_PREPROCESS_OPTIONS: UnstructuredPreprocessOption[] = [
+  'clean_invalid_content',
+  'detect_document_structure',
+  'merge_short_content',
+  'filter_low_quality',
+  'deduplicate_content',
+  'preserve_context',
+]
+
+const CHUNK_METHODS: Array<{ value: ChunkMethod; label: string }> = [
+  { value: 'semantic', label: '自动语义切分' },
+  { value: 'heading', label: '按标题和段落' },
+  { value: 'fixed', label: '按固定长度' },
+  { value: 'custom', label: '自定义分隔符' },
+]
+
+const UNSTRUCTURED_NUMBER_LIMITS = {
+  chunkSize: { min: 200, max: 2000 },
+  chunkOverlap: { min: 0, max: 500 },
+  minChunkSize: { min: 20, max: 500 },
+  qaPairsPerChunk: { min: 1, max: 3 },
+} as const
+
+type UnstructuredNumberField = keyof typeof UNSTRUCTURED_NUMBER_LIMITS
+
+const advancedChunkSettingsOpen = ref(props.options.chunkMethod === 'custom')
+
+const smartPreprocessEnabled = computed(() => SMART_PREPROCESS_OPTIONS.every(
+  (option) => props.options.preprocessOptions.includes(option),
+))
+
+const desensitizeEnabled = computed(() => props.options.preprocessOptions.includes('desensitize'))
+
+const preserveSpecialContentEnabled = computed(() => (
+  props.options.preserveTables
+  && props.options.preserveCodeBlocks
+  && props.options.preserveLists
+))
+
+watch(() => props.options.chunkMethod, (method) => {
+  if (method === 'custom') advancedChunkSettingsOpen.value = true
+})
+
+watch(() => props.chunkValidationMessage, (message) => {
+  if (message) advancedChunkSettingsOpen.value = true
+})
+
+function updateField<K extends keyof UnstructuredProcessOptions>(
+  field: K,
+  value: UnstructuredProcessOptions[K],
+) {
+  emit('update:options', { ...props.options, [field]: value })
+}
+
+function updateGenerationOptions(value: GenerationControlOptions) {
+  emit('update:options', { ...props.options, ...value })
+}
+
+function updateSmartPreprocess(value: string | number | boolean) {
+  const enabled = Boolean(value)
+  const remainingOptions = props.options.preprocessOptions.filter(
+    (option) => !SMART_PREPROCESS_OPTIONS.includes(option),
+  )
+  updateField(
+    'preprocessOptions',
+    enabled ? [...SMART_PREPROCESS_OPTIONS, ...remainingOptions] : remainingOptions,
+  )
+}
+
+function updateDesensitize(value: string | number | boolean) {
+  const preprocessOptions: UnstructuredPreprocessOption[] = props.options.preprocessOptions.filter(
+    (option) => option !== 'desensitize',
+  )
+  if (Boolean(value)) preprocessOptions.push('desensitize')
+  updateField('preprocessOptions', preprocessOptions)
+}
+
+function updateSpecialContentProtection(value: string | number | boolean) {
+  const enabled = Boolean(value)
+  emit('update:options', {
+    ...props.options,
+    preserveTables: enabled,
+    preserveCodeBlocks: enabled,
+    preserveLists: enabled,
+  })
+}
+
+function updateUnstructuredNumber(field: UnstructuredNumberField, value: number | undefined) {
+  const limits = UNSTRUCTURED_NUMBER_LIMITS[field]
+  const parsedValue = Number(value)
+  const nextValue = Number.isFinite(parsedValue) ? parsedValue : limits.min
+  updateField(field, Math.min(limits.max, Math.max(limits.min, nextValue)))
+}
+
+function revealValidation() {
+  advancedChunkSettingsOpen.value = true
+}
+
+defineExpose({ revealValidation })
+</script>
+
+<template>
+  <div class="form-section unstructured-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>预处理选项</h3>
+        <p>默认使用推荐策略，只需决定是否需要脱敏</p>
+      </div>
+    </div>
+    <div class="generation-option-list compact-option-list">
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>智能预处理</strong>
+          <small>自动完成内容清理、结构解析、短段合并、质量过滤、去重及上下文保留</small>
+        </div>
+        <el-switch
+          :model-value="smartPreprocessEnabled"
+          aria-label="智能预处理"
+          inline-prompt
+          active-text="开"
+          inactive-text="关"
+          @update:model-value="updateSmartPreprocess"
+        />
+      </div>
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>敏感信息脱敏</strong>
+          <small>处理姓名、手机号、邮箱和证件号等信息</small>
+        </div>
+        <el-switch
+          :model-value="desensitizeEnabled"
+          aria-label="敏感信息脱敏"
+          inline-prompt
+          active-text="开"
+          inactive-text="关"
+          @update:model-value="updateDesensitize"
+        />
+      </div>
+    </div>
+  </div>
+
+  <div class="form-section chunk-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>切分选项</h3>
+        <p>以语义完整为优先，将长文档拆成可独立生成问答的内容块</p>
+      </div>
+    </div>
+
+    <div class="chunk-settings-grid is-basic">
+      <label class="config-field">
+        <span class="config-field-label">切分方式</span>
+        <el-select
+          :model-value="options.chunkMethod"
+          aria-label="切分方式"
+          @update:model-value="updateField('chunkMethod', $event)"
+        >
+          <el-option
+            v-for="method in CHUNK_METHODS"
+            :key="method.value"
+            :label="method.label"
+            :value="method.value"
+          />
+        </el-select>
+        <small>推荐使用自动语义切分，在长度限制内优先保留完整句段</small>
+      </label>
+
+      <label class="config-field">
+        <span class="config-field-label">切片长度</span>
+        <span class="unit-input">
+          <el-input-number
+            :model-value="options.chunkSize"
+            :min="200"
+            :max="2000"
+            :step="50"
+            :precision="0"
+            controls-position="right"
+            aria-label="切片长度"
+            @update:model-value="updateUnstructuredNumber('chunkSize', $event)"
+          />
+          <span>Token</span>
+        </span>
+        <small>单个切片的目标上限，默认 800 Token</small>
+      </label>
+
+      <label class="config-field">
+        <span class="config-field-label">重叠长度</span>
+        <span class="unit-input">
+          <el-input-number
+            :model-value="options.chunkOverlap"
+            :min="0"
+            :max="500"
+            :step="10"
+            :precision="0"
+            controls-position="right"
+            aria-label="重叠长度"
+            @update:model-value="updateUnstructuredNumber('chunkOverlap', $event)"
+          />
+          <span>Token</span>
+        </span>
+        <small>在相邻切片中最多重复保留的上下文，默认 100 Token</small>
+      </label>
+    </div>
+
+    <p class="chunk-estimation-note">
+      Token 数为轻量估算值，实际长度以训练使用的模型分词器为准。
+    </p>
+
+    <p v-if="chunkValidationMessage" class="option-validation-message" role="alert">
+      {{ chunkValidationMessage }}
+    </p>
+
+    <button
+      type="button"
+      class="advanced-settings-toggle"
+      :aria-expanded="advancedChunkSettingsOpen"
+      aria-controls="unstructured-advanced-settings"
+      @click="advancedChunkSettingsOpen = !advancedChunkSettingsOpen"
+    >
+      <span>
+        <strong>高级设置</strong>
+        <small>最小切片长度、自定义分隔符和特殊内容保护</small>
+      </span>
+      <i class="fa" :class="advancedChunkSettingsOpen ? 'fa-chevron-up' : 'fa-chevron-down'" />
+    </button>
+
+    <div
+      v-if="advancedChunkSettingsOpen"
+      id="unstructured-advanced-settings"
+      class="advanced-settings-panel"
+    >
+      <div class="advanced-settings-grid">
+        <label class="config-field">
+          <span class="config-field-label">最小切片长度</span>
+          <span class="unit-input">
+            <el-input-number
+              :model-value="options.minChunkSize"
+              :min="20"
+              :max="500"
+              :step="10"
+              :precision="0"
+              controls-position="right"
+              aria-label="最小切片长度"
+              @update:model-value="updateUnstructuredNumber('minChunkSize', $event)"
+            />
+            <span>Token</span>
+          </span>
+          <small>过短的尾部内容会尽量并入前一个切片</small>
+        </label>
+
+        <label v-if="options.chunkMethod === 'custom'" class="config-field">
+          <span class="config-field-label">自定义分隔符</span>
+          <el-input
+            :model-value="options.customDelimiter"
+            maxlength="40"
+            show-word-limit
+            placeholder="例如：--- 或 ###"
+            aria-label="自定义分隔符"
+            @update:model-value="updateField('customDelimiter', $event)"
+          />
+          <small>系统会优先在分隔符位置结束当前切片</small>
+        </label>
+      </div>
+
+      <div class="generation-option-row advanced-protection-row">
+        <div class="generation-option-copy">
+          <strong>保护表格、代码和列表</strong>
+          <small>避免切分点破坏特殊内容块的完整性</small>
+        </div>
+        <el-switch
+          :model-value="preserveSpecialContentEnabled"
+          aria-label="保护表格、代码和列表"
+          inline-prompt
+          active-text="开"
+          inactive-text="关"
+          @update:model-value="updateSpecialContentProtection"
+        />
+      </div>
+    </div>
+  </div>
+
+  <div class="form-section generation-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>生成选项</h3>
+        <p>只保留会直接影响输出的核心设置</p>
+      </div>
+    </div>
+    <div class="generation-option-list">
+      <GenerationOptionsPanel
+        :options="options"
+        :models="generationModels"
+        section="quality"
+        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        @update:options="updateGenerationOptions"
+      />
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>语义丰富表达</strong>
+          <small>使用大模型将问答表述得更自然、柔和</small>
+        </div>
+        <el-switch
+          :model-value="options.semanticEnrichment"
+          aria-label="非结构化语义丰富表达"
+          inline-prompt
+          active-text="开"
+          inactive-text="关"
+          @update:model-value="updateField('semanticEnrichment', Boolean($event))"
+        />
+      </div>
+
+      <div class="generation-option-row">
+        <div class="generation-option-copy">
+          <strong>每个切片生成数量</strong>
+          <small>每个内容切片最多生成 3 个不同角度的问答对</small>
+        </div>
+        <el-input-number
+          :model-value="options.qaPairsPerChunk"
+          :min="1"
+          :max="3"
+          :step="1"
+          :precision="0"
+          controls-position="right"
+          aria-label="每个切片生成数量"
+          @update:model-value="updateUnstructuredNumber('qaPairsPerChunk', $event)"
+        />
+      </div>
+
+      <DatasetSplitEditor
+        :model-value="options.datasetSplit"
+        aria-label-prefix="非结构化"
+        @update:model-value="updateField('datasetSplit', $event)"
+      />
+    </div>
+  </div>
+
+  <div class="form-section model-options-section">
+    <div class="section-title-row">
+      <div>
+        <h3>大模型</h3>
+        <p>选择数据生成模型，并设置模型输出内容的要求</p>
+      </div>
+    </div>
+    <div class="generation-option-list">
+      <GenerationOptionsPanel
+        :options="options"
+        :models="generationModels"
+        section="model"
+        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        @update:options="updateGenerationOptions"
+      />
+    </div>
+  </div>
+</template>
+
+<style scoped lang="scss">
+.form-section {
+  padding: 0 0 26px;
+  margin-bottom: 26px;
+  border-bottom: 1px solid #edf0f5;
+
+  &:last-child {
+    margin-bottom: 0;
+    border-bottom: 0;
+  }
+
+  h3 {
+    margin: 0 0 16px;
+    color: #2f3747;
+    font-size: 15px;
+    font-weight: 650;
+  }
+}
+
+.section-title-row {
+  display: flex;
+  align-items: flex-start;
+  justify-content: space-between;
+  gap: 16px;
+
+  h3 {
+    margin-bottom: 5px;
+  }
+
+  p {
+    margin: 0;
+    color: #8a93a3;
+    font-size: 12px;
+  }
+}
+
+.generation-option-list {
+  display: grid;
+  gap: 12px;
+  margin-top: 16px;
+}
+
+.compact-option-list {
+  gap: 8px;
+
+  .generation-option-row {
+    min-height: 58px;
+  }
+}
+
+.generation-option-row {
+  display: flex;
+  min-height: 64px;
+  align-items: center;
+  justify-content: space-between;
+  gap: 24px;
+  padding: 12px 14px;
+  box-sizing: border-box;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+}
+
+.generation-option-copy {
+  display: grid;
+  gap: 4px;
+
+  strong {
+    color: #344054;
+    font-size: 13px;
+    font-weight: 600;
+  }
+
+  small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
+  }
+}
+
+.chunk-settings-grid {
+  display: grid;
+  grid-template-columns: repeat(3, minmax(0, 1fr));
+  gap: 12px;
+  margin-top: 16px;
+}
+
+.config-field {
+  display: grid;
+  align-content: start;
+  gap: 8px;
+  min-width: 0;
+  padding: 14px;
+  color: #344054;
+  background: #fff;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+
+  > small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
+  }
+
+  :deep(.el-select),
+  :deep(.el-input) {
+    width: 100%;
+  }
+}
+
+.config-field-label {
+  font-size: 13px;
+  font-weight: 600;
+}
+
+.unit-input {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  color: #7b8495;
+  font-size: 12px;
+
+  :deep(.el-input-number) {
+    width: 100%;
+  }
+}
+
+.option-validation-message {
+  display: block;
+  margin: 8px 0 0;
+  color: #dc2626;
+  font-size: 12px;
+  line-height: 1.5;
+}
+
+.chunk-estimation-note {
+  margin: 9px 0 0;
+  color: #8a93a3;
+  font-size: 12px;
+  line-height: 1.5;
+}
+
+.advanced-settings-toggle {
+  display: flex;
+  width: 100%;
+  min-height: 48px;
+  align-items: center;
+  justify-content: space-between;
+  gap: 16px;
+  margin-top: 12px;
+  padding: 10px 14px;
+  color: #344054;
+  text-align: left;
+  background: #fff;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+  cursor: pointer;
+  transition: border-color 0.18s ease, background-color 0.18s ease;
+
+  &:hover {
+    background: #f8f7ff;
+    border-color: #b7b2f7;
+  }
+
+  &:focus-visible {
+    outline: 2px solid #5b50f2;
+    outline-offset: 2px;
+  }
+
+  > span {
+    display: grid;
+    gap: 3px;
+  }
+
+  strong {
+    font-size: 13px;
+    font-weight: 600;
+  }
+
+  small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
+  }
+
+  i {
+    flex: 0 0 auto;
+    color: #7b8495;
+    font-size: 12px;
+  }
+}
+
+.advanced-settings-panel {
+  display: grid;
+  gap: 12px;
+  margin-top: 8px;
+  padding: 12px;
+  background: #fff;
+  border: 1px solid #e2e5ec;
+  border-radius: 8px;
+}
+
+.advanced-settings-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 12px;
+}
+
+@media (max-width: 900px) {
+  .chunk-settings-grid,
+  .advanced-settings-grid {
+    grid-template-columns: minmax(0, 1fr);
+  }
+
+  .generation-option-row {
+    align-items: flex-start;
+    flex-direction: column;
+  }
+
+  .compact-option-list .generation-option-row,
+  .advanced-protection-row {
+    align-items: center;
+    flex-direction: row;
+  }
+}
+</style>

frontend/src/views/data-process/create/data-process-create.scss

@@ -0,0 +1,193 @@
+.create-wizard-layout {
+  display: flex;
+  flex-direction: column;
+  height: 100%;
+  border: 1px solid #eef0f5;
+  border-radius: 8px;
+  background: #fff;
+}
+
+.wizard-main {
+  flex: 1;
+  overflow-y: auto;
+  padding: 32px;
+
+  &::-webkit-scrollbar {
+    width: 6px;
+  }
+
+  &::-webkit-scrollbar-thumb {
+    background: #cbd5e1;
+    border-radius: 3px;
+  }
+}
+
+.wizard-main-inner {
+  min-height: 100%;
+  box-sizing: border-box;
+}
+
+.wizard-steps-container {
+  margin-bottom: 32px;
+  padding-bottom: 24px;
+  border-bottom: 1px dashed #e2e8f0;
+}
+
+.custom-wizard-steps {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  max-width: 860px;
+  margin: 0 auto;
+  padding: 0 20px;
+}
+
+.step-item {
+  display: flex;
+  align-items: center;
+  flex: 1;
+
+  &:first-child {
+    flex: 0;
+  }
+}
+
+.step-connector {
+  flex: 1;
+  height: 2px;
+  margin: 0 16px;
+  background-color: #e2e8f0;
+  transition: background-color 0.3s;
+}
+
+.step-item.is-completed .step-connector,
+.step-item.is-active .step-connector {
+  background-color: #5146e5;
+}
+
+.step-node {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+}
+
+.step-icon {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  width: 28px;
+  height: 28px;
+  color: #64748b;
+  font-size: 13px;
+  font-weight: 650;
+  background-color: #fff;
+  border: 2px solid #cbd5e1;
+  border-radius: 50%;
+  transition: all 0.3s;
+}
+
+.step-title {
+  color: #64748b;
+  font-size: 14px;
+  font-weight: 600;
+  white-space: nowrap;
+}
+
+.step-item.is-active {
+  .step-icon {
+    color: #5146e5;
+    background-color: #eef2ff;
+    border-color: #5146e5;
+  }
+
+  .step-title {
+    color: #1e293b;
+  }
+}
+
+.step-item.is-completed {
+  .step-icon {
+    color: #fff;
+    background-color: #5146e5;
+    border-color: #5146e5;
+  }
+
+  .step-title {
+    color: #1e293b;
+  }
+}
+
+.wizard-content {
+  min-height: 400px;
+}
+
+.wizard-footer {
+  display: grid;
+  flex-shrink: 0;
+  grid-template-columns: 1fr auto 1fr;
+  align-items: center;
+  height: 64px;
+  padding: 0 32px;
+  background: #fff;
+  border-top: 1px solid #e2e8f0;
+  box-shadow: 0 -4px 6px -1px rgba(0, 0, 0, 0.02);
+
+  .footer-left {
+    justify-self: start;
+  }
+
+  .footer-center {
+    justify-self: center;
+  }
+
+  .footer-right {
+    justify-self: end;
+  }
+}
+
+.wizard-primary-action {
+  min-width: 160px;
+}
+
+@media (max-width: 760px) {
+  .wizard-main {
+    padding: 24px 20px;
+  }
+
+  .custom-wizard-steps {
+    padding: 0;
+  }
+
+  .step-connector {
+    margin: 0 8px;
+  }
+
+  .step-node {
+    gap: 6px;
+  }
+
+  .step-title {
+    max-width: 72px;
+    font-size: 12px;
+    line-height: 1.35;
+    white-space: normal;
+  }
+
+  .wizard-footer {
+    padding: 0 20px;
+  }
+}
+
+@media (max-width: 600px) {
+  .step-title {
+    display: none;
+  }
+
+  .step-connector {
+    margin: 0 6px;
+  }
+
+  .wizard-primary-action {
+    min-width: 0;
+  }
+}

frontend/src/views/data-process/create/dataProcessCreateState.ts

@@ -0,0 +1,54 @@
+import type { StructuredProcessOptions, UnstructuredProcessOptions } from './types'
+
+export const DEFAULT_GENERATION_PROMPT = '你是一名专业的数据生成助手。请根据输入内容生成准确、完整、可直接用于模型训练的问答数据。仅输出符合目标格式的内容，答案应事实清晰、语言自然，不要添加分析过程、说明或无关内容。'
+
+export function createDefaultStructuredOptions(): StructuredProcessOptions {
+  return {
+    preprocessOptions: ['clean_invalid', 'detect_structure', 'deduplicate', 'normalize_format'],
+    semanticEnrichment: false,
+    qaPairsPerRow: 1,
+    datasetSplit: { train: 80, validation: 10, test: 10 },
+    generationModelId: '',
+    generationPrompt: DEFAULT_GENERATION_PROMPT,
+    temperature: 0.7,
+    maxTokens: 1024,
+    jsonMode: false,
+    qualityFilterEnabled: false,
+    filterLowQuality: true,
+    filterShortContent: true,
+    minOutputLength: 20,
+  }
+}
+
+export function createDefaultUnstructuredOptions(): UnstructuredProcessOptions {
+  return {
+    preprocessOptions: [
+      'clean_invalid_content',
+      'detect_document_structure',
+      'merge_short_content',
+      'filter_low_quality',
+      'deduplicate_content',
+      'preserve_context',
+    ],
+    chunkMethod: 'semantic',
+    chunkSize: 800,
+    chunkOverlap: 100,
+    minChunkSize: 100,
+    customDelimiter: '',
+    preserveTables: true,
+    preserveCodeBlocks: true,
+    preserveLists: true,
+    semanticEnrichment: false,
+    qaPairsPerChunk: 1,
+    datasetSplit: { train: 80, validation: 10, test: 10 },
+    generationModelId: '',
+    generationPrompt: DEFAULT_GENERATION_PROMPT,
+    temperature: 0.7,
+    maxTokens: 1024,
+    jsonMode: false,
+    qualityFilterEnabled: false,
+    filterLowQuality: true,
+    filterShortContent: true,
+    minOutputLength: 20,
+  }
+}

frontend/src/views/data-process/create/types.ts

@@ -71,6 +71,14 @@ export interface ExternalDataSource {
   limit: number
 }
 
+export interface UploadedDataFile {
+  uid: string | number
+  name: string
+  size: number
+  count: number
+  content: string
+}
+
 export interface SourceLine {
   number: number
   content: string

frontend/src/views/data-process/create/useDataProcessDraft.ts

@@ -0,0 +1,147 @@
+import { nextTick, type Reactive, type Ref } from 'vue'
+import { ElMessage } from 'element-plus'
+import type {
+  ExternalDataSource,
+  ProcessType,
+  StepId,
+  StructuredProcessOptions,
+  UnstructuredProcessOptions,
+} from './types'
+
+export const DATA_PROCESS_DRAFT_STORAGE_KEY = 'yg-data-process-create-draft'
+export const DATA_PROCESS_DRAFT_SCHEMA_VERSION = 6
+
+interface DraftSnapshot {
+  schemaVersion?: number
+  currentStepId?: StepId
+  task?: { name?: string; description?: string }
+  processType?: ProcessType
+  structuredOptions?: Partial<StructuredProcessOptions>
+  unstructuredOptions?: Partial<UnstructuredProcessOptions>
+  externalSource?: Partial<ExternalDataSource>
+}
+
+interface DraftBindings {
+  currentStepId: Readonly<Ref<StepId>>
+  task: Reactive<{ name: string; description: string }>
+  processType: Ref<ProcessType>
+  structuredOptions: Ref<StructuredProcessOptions>
+  unstructuredOptions: Ref<UnstructuredProcessOptions>
+  externalSource: Reactive<ExternalDataSource>
+  restoringDraft: Ref<boolean>
+  dirty: Ref<boolean>
+  goToStep: (stepId: StepId) => void
+}
+
+function sanitizeExternalSource(source: Partial<ExternalDataSource>) {
+  return {
+    type: typeof source.type === 'string' ? source.type : 'mysql',
+    url: typeof source.url === 'string' ? source.url : '',
+    authMode: typeof source.authMode === 'string' ? source.authMode : 'none',
+    username: typeof source.username === 'string' ? source.username : '',
+    limit: Number.isFinite(source.limit) ? Number(source.limit) : 1000,
+  }
+}
+
+function isDraftSnapshot(value: unknown): value is DraftSnapshot {
+  return Boolean(value && typeof value === 'object')
+}
+
+/**
+ * 只持久化可重建的配置。密码、令牌、文件正文、预览与结果都不进入 localStorage。
+ */
+export function useDataProcessDraft(bindings: DraftBindings) {
+  function draftSnapshot(): DraftSnapshot {
+    return {
+      schemaVersion: DATA_PROCESS_DRAFT_SCHEMA_VERSION,
+      currentStepId: bindings.currentStepId.value,
+      task: { ...bindings.task },
+      processType: bindings.processType.value,
+      structuredOptions: {
+        ...bindings.structuredOptions.value,
+        preprocessOptions: [...bindings.structuredOptions.value.preprocessOptions],
+        datasetSplit: { ...bindings.structuredOptions.value.datasetSplit },
+      },
+      unstructuredOptions: {
+        ...bindings.unstructuredOptions.value,
+        preprocessOptions: [...bindings.unstructuredOptions.value.preprocessOptions],
+        datasetSplit: { ...bindings.unstructuredOptions.value.datasetSplit },
+      },
+      externalSource: sanitizeExternalSource(bindings.externalSource),
+    }
+  }
+
+  function writeDraft(showWarning: boolean) {
+    try {
+      localStorage.setItem(DATA_PROCESS_DRAFT_STORAGE_KEY, JSON.stringify(draftSnapshot()))
+    } catch {
+      if (showWarning) ElMessage.warning('草稿保存失败，请检查浏览器存储空间')
+    }
+  }
+
+  function persistDraft() {
+    if (!bindings.restoringDraft.value) writeDraft(true)
+  }
+
+  function restoreDraft() {
+    try {
+      const raw = localStorage.getItem(DATA_PROCESS_DRAFT_STORAGE_KEY)
+      if (!raw) return
+      const snapshot: unknown = JSON.parse(raw)
+      if (!isDraftSnapshot(snapshot)) return
+
+      bindings.restoringDraft.value = true
+      bindings.goToStep('create')
+      bindings.task.name = snapshot.task?.name || ''
+      bindings.task.description = snapshot.task?.description || ''
+      bindings.processType.value = snapshot.processType === 'structured' || snapshot.processType === 'external'
+        ? snapshot.processType
+        : 'unstructured'
+
+      if (snapshot.structuredOptions) {
+        bindings.structuredOptions.value = {
+          ...bindings.structuredOptions.value,
+          ...snapshot.structuredOptions,
+          preprocessOptions: Array.isArray(snapshot.structuredOptions.preprocessOptions)
+            ? snapshot.structuredOptions.preprocessOptions
+            : bindings.structuredOptions.value.preprocessOptions,
+          datasetSplit: {
+            ...bindings.structuredOptions.value.datasetSplit,
+            ...snapshot.structuredOptions.datasetSplit,
+          },
+        }
+      }
+
+      if (snapshot.unstructuredOptions) {
+        bindings.unstructuredOptions.value = {
+          ...bindings.unstructuredOptions.value,
+          ...snapshot.unstructuredOptions,
+          preprocessOptions: Array.isArray(snapshot.unstructuredOptions.preprocessOptions)
+            ? snapshot.unstructuredOptions.preprocessOptions
+            : bindings.unstructuredOptions.value.preprocessOptions,
+          datasetSplit: {
+            ...bindings.unstructuredOptions.value.datasetSplit,
+            ...snapshot.unstructuredOptions.datasetSplit,
+          },
+        }
+      }
+
+      Object.assign(bindings.externalSource, sanitizeExternalSource(snapshot.externalSource || {}), {
+        password: '',
+        token: '',
+      })
+      bindings.dirty.value = false
+
+      nextTick(() => {
+        bindings.restoringDraft.value = false
+        // 立即覆盖 v5 及更早草稿，清除其中可能存在的敏感值和大段正文。
+        writeDraft(false)
+      })
+      ElMessage.info('已恢复上次的任务配置，请重新上传或拉取源数据')
+    } catch {
+      localStorage.removeItem(DATA_PROCESS_DRAFT_STORAGE_KEY)
+    }
+  }
+
+  return { draftSnapshot, persistDraft, restoreDraft }
+}

frontend/src/views/data-process/create/useDataProcessGeneration.ts

@@ -0,0 +1,127 @@
+import { reactive, ref, type Ref } from 'vue'
+import { ElMessage } from 'element-plus'
+import { createResults } from './previewModel'
+import type {
+  GenerationState,
+  PreviewItem,
+  ProcessType,
+  ResultItem,
+  StructuredProcessOptions,
+  UnstructuredProcessOptions,
+} from './types'
+
+interface GenerationBindings {
+  previewItems: Ref<PreviewItem[]>
+  processType: Ref<ProcessType>
+  structuredOptions: Ref<StructuredProcessOptions>
+  unstructuredOptions: Ref<UnstructuredProcessOptions>
+  dirty: Ref<boolean>
+}
+
+export function useDataProcessGeneration(bindings: GenerationBindings) {
+  const results = ref<ResultItem[]>([])
+  const selectedResultId = ref<string | null>(null)
+  const generation = reactive<GenerationState>({
+    status: 'idle',
+    progress: 0,
+    message: '确认摘要后即可开始生成，过程中可查看实时进度。',
+  })
+  let generationTimer: ReturnType<typeof setInterval> | null = null
+
+  function stopGenerationTimer() {
+    if (generationTimer) clearInterval(generationTimer)
+    generationTimer = null
+  }
+
+  function resetDownstream() {
+    stopGenerationTimer()
+    generation.status = 'idle'
+    generation.progress = 0
+    generation.message = '确认摘要后即可开始生成，过程中可查看实时进度。'
+    results.value = []
+    selectedResultId.value = null
+  }
+
+  function startGeneration() {
+    stopGenerationTimer()
+    generation.status = 'running'
+    generation.progress = 0
+    generation.message = '正在应用预览修改并生成标准化结果，请稍候。'
+
+    generationTimer = setInterval(() => {
+      generation.progress = Math.min(100, generation.progress + 8)
+      if (generation.progress < 100) return
+
+      stopGenerationTimer()
+      generation.status = 'success'
+      results.value = createResults(
+        bindings.previewItems.value,
+        bindings.processType.value === 'structured'
+          ? bindings.structuredOptions.value
+          : bindings.processType.value === 'unstructured'
+            ? bindings.unstructuredOptions.value
+            : undefined,
+      )
+      generation.message = `已完成 ${results.value.length.toLocaleString()} 条数据处理，可进入结果页检查。`
+      selectedResultId.value = results.value[0]?.id ?? null
+      bindings.dirty.value = true
+      ElMessage.success('数据处理完成')
+    }, 180)
+  }
+
+  function stopGeneration() {
+    stopGenerationTimer()
+    generation.status = 'failed'
+    generation.message = '任务已停止，预览修改仍然保留，可以重新生成。'
+  }
+
+  function updateResultField(id: string, field: 'instruction' | 'input' | 'output', value: string) {
+    const item = results.value.find((entry) => entry.id === id)
+    if (!item) return
+    item[field] = value
+    const valid = item.instruction.trim() && item.output.trim()
+    item.error = valid ? undefined : 'Instruction 和 Output 不能为空'
+    const changed = item.instruction !== item.originalInstruction
+      || item.input !== item.originalInput
+      || item.output !== item.originalOutput
+    item.status = item.error ? 'invalid' : changed ? 'modified' : 'valid'
+    bindings.dirty.value = true
+  }
+
+  function restoreResult(id: string) {
+    const item = results.value.find((entry) => entry.id === id)
+    if (!item) return
+    item.instruction = item.originalInstruction
+    item.input = item.originalInput
+    item.output = item.originalOutput
+    item.error = undefined
+    item.status = 'valid'
+    bindings.dirty.value = true
+  }
+
+  function validateResults() {
+    let firstInvalidId: string | null = null
+    for (const item of results.value) {
+      if (!item.instruction.trim() || !item.output.trim()) {
+        item.error = 'Instruction 和 Output 不能为空'
+        item.status = 'invalid'
+        firstInvalidId ??= item.id
+      }
+    }
+    if (firstInvalidId) selectedResultId.value = firstInvalidId
+    return firstInvalidId == null
+  }
+
+  return {
+    generation,
+    results,
+    selectedResultId,
+    resetDownstream,
+    restoreResult,
+    startGeneration,
+    stopGeneration,
+    stopGenerationTimer,
+    updateResultField,
+    validateResults,
+  }
+}

frontend/src/views/dataset/DatasetPreviewView.vue

@@ -14,14 +14,13 @@ import {
 import { DATASET_TYPE_MAP, STORAGE_MAP } from '@/constants'
 import { parseDatasetRecords, updateDatasetRecord } from './datasetRecords'
 import type { DatasetRecord } from './datasetRecords'
+import DatasetVersionBar from './preview/DatasetVersionBar.vue'
+import DatasetRecordTable from './preview/DatasetRecordTable.vue'
+import DatasetRecordEditorDialog from './preview/DatasetRecordEditorDialog.vue'
+import DatasetRawPreview from './preview/DatasetRawPreview.vue'
+import type { EditField } from './preview/types'
 import type { DatasetFile, DatasetItem, DatasetVersion } from '@/types'
 
-interface EditField {
-  key: string
-  value: string
-  isJson: boolean
-}
-
 const route = useRoute()
 const datasetId = route.params.id as string
 
@@ -32,9 +31,6 @@ const selectedFileId = ref('')
 const previewContent = ref('')
 const baseVersionContent = ref('')
 const previewError = ref('')
-const searchText = ref('')
-const currentPage = ref(1)
-const pageSize = 10
 const editorVisible = ref(false)
 const editingRecord = ref<DatasetRecord | null>(null)
 const editFields = ref<EditField[]>([])
@@ -56,37 +52,9 @@ const totalLines = computed(() => previewContent.value ? previewContent.value.sp
 const recordResult = computed(() => parseDatasetRecords(previewContent.value, selectedFile.value?.name || ''))
 const isRecordFile = computed(() => recordResult.value.supported)
 const records = computed(() => recordResult.value.records)
-const invalidRecordCount = computed(() => records.value.filter((record) => record.kind === 'invalid').length)
-const allTableFieldKeys = computed(() => {
-  const keys = new Set<string>()
-  records.value.forEach((record) => {
-    if (record.kind !== 'object') return
-    Object.keys(record.value as Record<string, unknown>).forEach((key) => keys.add(key))
-  })
-  const preferred = ['instruction', 'question', 'prompt', 'input', 'context', 'output', 'answer', 'response']
-  return [...keys].sort((a, b) => {
-    const rank = (key: string) => {
-      const index = preferred.indexOf(key)
-      return index === -1 ? preferred.length : index
-    }
-    return rank(a) - rank(b)
-  })
-})
-const tableFieldKeys = computed(() => allTableFieldKeys.value.slice(0, 6))
-const hasExtraTableFields = computed(() => allTableFieldKeys.value.length > tableFieldKeys.value.length)
 const viewedVersion = computed(() => versions.value.find((item) => item.id === loadedVersionId.value))
 const isViewingActiveVersion = computed(() => Boolean(loadedVersionId.value) && loadedVersionId.value === activeVersionId.value)
 const nextVersionNumber = computed(() => Math.max(0, ...versions.value.map((item) => item.version)) + 1)
-const filteredRecords = computed(() => {
-  const query = searchText.value.trim().toLowerCase()
-  if (!query) return records.value
-  return records.value.filter((record) => JSON.stringify(record.value).toLowerCase().includes(query))
-})
-const pagedRecords = computed(() => {
-  const start = (currentPage.value - 1) * pageSize
-  return filteredRecords.value.slice(start, start + pageSize)
-})
-const rawPreviewLines = computed(() => previewContent.value ? previewContent.value.split('\n').slice(0, 100) : [])
 const currentDraft = computed(() => editingRecord.value?.kind === 'object'
   ? JSON.stringify(editFields.value)
   : rawDraft.value)
@@ -105,47 +73,10 @@ const createdAt = computed(() => dataset.value?.create_time
   ? new Date(dataset.value.create_time).toLocaleString('zh-CN')
   : '-')
 
-function formatVersionTime(value: string) {
-  return new Date(value).toLocaleString('zh-CN', { hour12: false })
-}
-
-function versionOptionLabel(version: DatasetVersion) {
-  const status = version.id === activeVersionId.value ? '当前版本' : '历史版本'
-  return `V${version.version} · ${status} · ${formatVersionTime(version.create_time)}`
-}
-
 function fileKey(file: DatasetFile) {
   return String(file.id || file.name)
 }
 
-function formatFieldValue(value: unknown) {
-  if (typeof value === 'string') return value || '（空）'
-  return JSON.stringify(value)
-}
-
-function asDatasetRecord(row: unknown) {
-  return row as DatasetRecord
-}
-
-function recordFieldValue(record: DatasetRecord, key: string) {
-  if (record.kind !== 'object') return record.kind === 'invalid' ? record.raw : formatFieldValue(record.value)
-  const data = record.value as Record<string, unknown>
-  return Object.prototype.hasOwnProperty.call(data, key) ? formatFieldValue(data[key]) : '-'
-}
-
-function extraFieldValue(record: DatasetRecord) {
-  if (record.kind !== 'object') return '-'
-  const data = record.value as Record<string, unknown>
-  return allTableFieldKeys.value
-    .filter((key) => !tableFieldKeys.value.includes(key) && Object.prototype.hasOwnProperty.call(data, key))
-    .map((key) => `${key}: ${formatFieldValue(data[key])}`)
-    .join('；') || '-'
-}
-
-function recordRowClassName({ row }: { row: unknown }) {
-  return asDatasetRecord(row).kind === 'invalid' ? 'record-table-row-invalid' : ''
-}
-
 async function loadDataset() {
   loading.value = true
   try {
@@ -174,6 +105,7 @@ async function loadVersions(file: DatasetFile) {
     loadedVersionId.value = result.active_version_id
     previewContent.value = activeContent.content
     baseVersionContent.value = activeContent.content
+    resetRecordEditor()
   } catch {
     if (requestId === versionRequestId) previewError.value = '版本内容加载失败，请稍后重试'
   } finally {
@@ -197,8 +129,7 @@ async function handleViewedVersionChange(nextVersionId: string) {
     previewContent.value = result.content
     baseVersionContent.value = result.content
     loadedVersionId.value = result.version.id
-    currentPage.value = 1
+    resetRecordEditor()
-    searchText.value = ''
   } catch {
     if (requestId === versionRequestId) selectedVersionId.value = previousVersionId
   } finally {
@@ -259,6 +190,19 @@ function openRecordEditor(record: DatasetRecord) {
   originalDraft.value = record.kind === 'object' ? JSON.stringify(editFields.value) : rawDraft.value
 }
 
+function resetRecordEditor() {
+  editorVisible.value = false
+  editingRecord.value = null
+  editFields.value = []
+  rawDraft.value = ''
+  originalDraft.value = ''
+}
+
+function updateEditField(index: number, value: string) {
+  const field = editFields.value[index]
+  if (field) field.value = value
+}
+
 async function confirmDiscardChanges(message = '当前存在尚未保存为版本的修改，放弃后将无法恢复。') {
   if (!hasUnsavedChanges.value) return true
   try {
@@ -287,18 +231,11 @@ async function confirmEditorDiscardChanges() {
   }
 }
 
-async function closeRecordEditor() {
+async function closeRecordEditor(done?: () => void) {
   if (savingRecord.value) return
   if (!await confirmEditorDiscardChanges()) return
-  editorVisible.value = false
+  resetRecordEditor()
-  editingRecord.value = null
+  done?.()
-}
-
-async function handleDialogBeforeClose(done: () => void) {
-  if (savingRecord.value) return
-  if (!await confirmEditorDiscardChanges()) return
-  editingRecord.value = null
-  done()
 }
 
 function buildNextRecordValue() {
@@ -330,9 +267,7 @@ async function saveRecord() {
       nextValue,
     )
     previewContent.value = nextContent
-    originalDraft.value = currentDraft.value
+    resetRecordEditor()
-    editorVisible.value = false
-    editingRecord.value = null
     ElMessage.success(`第 ${record.displayIndex} 条修改已暂存`)
   } finally {
     savingRecord.value = false
@@ -370,13 +305,6 @@ async function saveVersion() {
   }
 }
 
-function handleEditorKeydown(event: KeyboardEvent) {
-  if ((event.ctrlKey || event.metaKey) && event.key.toLowerCase() === 's') {
-    event.preventDefault()
-    saveRecord()
-  }
-}
-
 function handleBeforeUnload(event: BeforeUnloadEvent) {
   if (!hasUnsavedChanges.value) return
   event.preventDefault()
@@ -419,46 +347,19 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
       <div class="overview-item overview-item-wide"><span>创建时间</span><strong>{{ createdAt }}</strong></div>
     </section>
 
-    <section class="version-control-bar" aria-label="数据集版本控制">
+    <DatasetVersionBar
-      <div class="version-selector-group">
-        <label for="dataset-version-select">数据版本</label>
-        <el-select
-          id="dataset-version-select"
       v-model="selectedVersionId"
-          class="version-select"
+      :versions="versions"
-          :loading="versionLoading"
+      :viewed-version="viewedVersion"
-          :disabled="versionLoading || savingRecord"
+      :active-version-id="activeVersionId"
-          aria-label="选择要查看的数据集版本"
+      :is-viewing-active-version="isViewingActiveVersion"
+      :has-pending-version-changes="hasPendingVersionChanges"
+      :version-loading="versionLoading"
+      :saving-record="savingRecord"
+      :activating-version="activatingVersion"
       @change="handleViewedVersionChange"
-        >
+      @activate="activateViewedVersion"
-          <el-option
-            v-for="version in versions"
-            :key="version.id"
-            :label="versionOptionLabel(version)"
-            :value="version.id"
     />
-        </el-select>
-      </div>
-
-      <div class="version-status" role="status" aria-live="polite">
-        <template v-if="viewedVersion">
-          <el-tag v-if="isViewingActiveVersion" type="success">V{{ viewedVersion.version }} · 当前版本</el-tag>
-          <el-tag v-else type="info">V{{ viewedVersion.version }} · 历史版本（只读）</el-tag>
-          <el-tag v-if="hasPendingVersionChanges" type="warning">有待保存修改</el-tag>
-          <span>{{ viewedVersion.description || '无版本说明' }} · {{ formatVersionTime(viewedVersion.create_time) }}</span>
-        </template>
-      </div>
-
-      <el-button
-        v-if="viewedVersion && !isViewingActiveVersion"
-        type="primary"
-        :loading="activatingVersion"
-        @click="activateViewedVersion"
-      >
-        <i class="fa fa-check-circle" />
-        设为当前版本
-      </el-button>
-    </section>
 
     <el-alert
       v-if="viewedVersion && !isViewingActiveVersion"
@@ -508,78 +409,19 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
           </div>
           <div v-else-if="!previewContent" class="viewer-state"><i class="fa fa-file-o" aria-hidden="true" /><span>当前文件暂无可预览内容</span></div>
 
-          <div v-else-if="isRecordFile" class="records-viewer">
+          <DatasetRecordTable
-            <div class="records-toolbar">
+            v-else-if="isRecordFile"
-              <div><strong>样本数据</strong><span>共 {{ records.length.toLocaleString() }} 条</span><el-tag v-if="invalidRecordCount" size="small" type="danger">{{ invalidRecordCount }} 条格式错误</el-tag></div>
+            :records="records"
-              <el-input v-model="searchText" class="record-search" clearable placeholder="搜索样本内容" aria-label="搜索样本内容" @input="currentPage = 1">
+            :is-viewing-active-version="isViewingActiveVersion"
-                <template #prefix><i class="fa fa-search" /></template>
+            :version-id="loadedVersionId"
-              </el-input>
+            @edit="openRecordEditor"
-            </div>
+          />
 
-            <div v-if="!filteredRecords.length" class="viewer-state record-empty-state"><i class="fa fa-search" aria-hidden="true" /><span>没有找到匹配的样本</span></div>
+          <DatasetRawPreview
-            <div v-else class="record-table-shell" aria-label="数据样本表格">
+            v-else
-              <el-table
+            :content="previewContent"
-                :data="pagedRecords"
+            :file-name="selectedFile?.name"
-                row-key="sourceIndex"
+          />
-                stripe
-                border
-                class="record-table"
-                :row-class-name="recordRowClassName"
-              >
-                <el-table-column label="序号" width="76" align="center">
-                  <template #default="{ row }">
-                    <div class="table-index">
-                      <strong>{{ asDatasetRecord(row).displayIndex }}</strong>
-                      <small>第 {{ asDatasetRecord(row).sourceIndex + 1 }} 行</small>
-                    </div>
-                  </template>
-                </el-table-column>
-                <el-table-column v-for="fieldKey in tableFieldKeys" :key="fieldKey" :label="fieldKey" min-width="220">
-                  <template #default="{ row }">
-                    <el-tooltip :content="recordFieldValue(asDatasetRecord(row), fieldKey)" placement="top" :show-after="400">
-                      <div class="table-cell-text">{{ recordFieldValue(asDatasetRecord(row), fieldKey) }}</div>
-                    </el-tooltip>
-                  </template>
-                </el-table-column>
-                <el-table-column v-if="hasExtraTableFields" label="其他字段" min-width="240">
-                  <template #default="{ row }">
-                    <el-tooltip :content="extraFieldValue(asDatasetRecord(row))" placement="top" :show-after="400">
-                      <div class="table-cell-text">{{ extraFieldValue(asDatasetRecord(row)) }}</div>
-                    </el-tooltip>
-                  </template>
-                </el-table-column>
-                <el-table-column label="状态" width="96" align="center">
-                  <template #default="{ row }">
-                    <el-tag v-if="asDatasetRecord(row).kind === 'invalid'" size="small" type="danger">格式错误</el-tag>
-                    <el-tag v-else-if="asDatasetRecord(row).kind === 'primitive'" size="small" type="warning">非对象</el-tag>
-                    <el-tag v-else size="small" type="success">正常</el-tag>
-                  </template>
-                </el-table-column>
-                <el-table-column label="操作" width="96" align="center" fixed="right">
-                  <template #default="{ row }">
-                    <el-button
-                      type="primary"
-                      link
-                      :disabled="!isViewingActiveVersion"
-                      :title="isViewingActiveVersion ? '' : '历史版本为只读，请先设为当前版本'"
-                      :aria-label="`${asDatasetRecord(row).kind === 'invalid' ? '修复' : '编辑'}第 ${asDatasetRecord(row).displayIndex} 条数据`"
-                      @click="openRecordEditor(asDatasetRecord(row))"
-                    >
-                      <i :class="asDatasetRecord(row).kind === 'invalid' ? 'fa fa-wrench' : 'fa fa-pencil'" />
-                      {{ asDatasetRecord(row).kind === 'invalid' ? '修复' : '编辑' }}
-                    </el-button>
-                  </template>
-                </el-table-column>
-              </el-table>
-            </div>
-            <div v-if="filteredRecords.length > pageSize" class="records-pagination">
-              <el-pagination v-model:current-page="currentPage" :page-size="pageSize" :total="filteredRecords.length" layout="prev, pager, next" background />
-            </div>
-          </div>
-
-          <div v-else class="code-viewer" tabindex="0" :aria-label="`${selectedFile?.name || ''} 文件内容`">
-            <div v-for="(line, index) in rawPreviewLines" :key="index" class="code-line"><span class="line-number" aria-hidden="true">{{ index + 1 }}</span><code>{{ line || ' ' }}</code></div>
-          </div>
         </div>
 
         <div class="viewer-footer"><span v-if="isRecordFile"><i class="fa fa-list-alt" aria-hidden="true" /> 逐条查看与编辑</span><span v-else><i class="fa fa-eye" aria-hidden="true" /> 文本预览</span><span>UTF-8</span></div>
@@ -590,30 +432,19 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
       <span class="empty-icon" aria-hidden="true"><i class="fa fa-folder-open-o" /></span><h2>暂无数据文件</h2><p>该数据集还没有可供预览的文件。</p>
     </section>
 
-    <el-dialog v-model="editorVisible" class="record-editor-dialog" width="min(720px, 92vw)" :title="editingRecord ? `编辑第 ${editingRecord.displayIndex} 条数据` : '编辑数据'" :close-on-click-modal="false" :close-on-press-escape="!savingRecord" :show-close="!savingRecord" :before-close="handleDialogBeforeClose" destroy-on-close>
+    <DatasetRecordEditorDialog
-      <div v-if="editingRecord" class="record-editor" @keydown="handleEditorKeydown">
+      :model-value="editorVisible"
-        <div class="editor-hint">
+      :record="editingRecord"
-          <i class="fa fa-info-circle" aria-hidden="true" />
+      :edit-fields="editFields"
-          当前修改会先暂存；完成多条修改后，点击文件工具栏中的“保存版本”创建 V{{ nextVersionNumber }}。
+      :raw-draft="rawDraft"
-        </div>
+      :saving-record="savingRecord"
-        <el-form v-if="editingRecord.kind === 'object'" label-position="top">
+      :has-draft-changes="hasEditorDraftChanges"
-          <el-form-item v-for="field in editFields" :key="field.key" :label="field.key">
+      :next-version-number="nextVersionNumber"
-            <el-input v-model="field.value" type="textarea" :rows="field.isJson ? 5 : 3" :aria-label="`编辑字段 ${field.key}`" :placeholder="field.isJson ? '请输入合法 JSON' : `请输入 ${field.key}`" />
+      @close="closeRecordEditor"
-            <small v-if="field.isJson" class="field-helper">该字段为对象、数组或数值，请保持合法 JSON 格式。</small>
+      @save="saveRecord"
-          </el-form-item>
+      @update:field="updateEditField"
-        </el-form>
+      @update:raw-draft="rawDraft = $event"
-        <div v-else class="raw-record-editor">
+    />
-          <label for="raw-record-draft">{{ editingRecord.kind === 'invalid' ? '修复 JSON 内容' : 'JSON 内容' }}</label>
-          <el-input id="raw-record-draft" v-model="rawDraft" type="textarea" :rows="10" aria-label="编辑当前条 JSON 内容" placeholder="请输入合法 JSON" />
-        </div>
-      </div>
-      <template #footer>
-        <div class="record-editor-actions">
-          <span>{{ hasEditorDraftChanges ? '有未暂存修改' : '尚未修改' }} · Ctrl/⌘ + S 暂存</span>
-          <div><el-button :disabled="savingRecord" @click="closeRecordEditor">取消</el-button><el-button type="primary" :loading="savingRecord" :disabled="!hasEditorDraftChanges" @click="saveRecord">暂存修改</el-button></div>
-        </div>
-      </template>
-    </el-dialog>
   </PageCard>
 </template>
 
@@ -715,61 +546,6 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
   white-space: nowrap;
 }
 
-.version-control-bar {
-  display: flex;
-  align-items: center;
-  gap: 16px;
-  min-height: 64px;
-  margin-bottom: 12px;
-  padding: 10px 14px;
-  box-sizing: border-box;
-  border: 1px solid #dfe5ee;
-  border-radius: 10px;
-  background: #fff;
-}
-
-.version-selector-group {
-  display: flex;
-  align-items: center;
-  gap: 10px;
-  flex: 0 0 auto;
-}
-
-.version-selector-group label {
-  color: #344054;
-  font-size: 13px;
-  font-weight: 600;
-}
-
-.version-select {
-  width: 300px;
-}
-
-.version-status {
-  display: flex;
-  align-items: center;
-  flex: 1;
-  gap: 10px;
-  min-width: 0;
-}
-
-.version-status > span {
-  overflow: hidden;
-  color: #7c8799;
-  font-size: 12px;
-  text-overflow: ellipsis;
-  white-space: nowrap;
-}
-
-.version-control-bar > .el-button {
-  min-height: 40px;
-  flex: 0 0 auto;
-}
-
-.version-control-bar > .el-button i {
-  margin-right: 6px;
-}
-
 .historical-version-alert {
   margin-bottom: 12px;
 }
@@ -868,153 +644,6 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
   background: #fbfcfe;
 }
 
-.records-viewer {
-  min-height: 468px;
-  padding: 16px;
-  box-sizing: border-box;
-}
-
-.records-toolbar {
-  display: flex;
-  align-items: center;
-  justify-content: space-between;
-  gap: 16px;
-  margin-bottom: 14px;
-}
-
-.records-toolbar > div {
-  display: flex;
-  align-items: center;
-  gap: 8px;
-  min-width: 0;
-}
-
-.records-toolbar strong {
-  color: #253047;
-  font-size: 14px;
-}
-
-.records-toolbar span {
-  color: #8a94a6;
-  font-size: 12px;
-}
-
-.record-search {
-  width: 240px;
-}
-
-.record-table-shell {
-  overflow: hidden;
-  border: 1px solid #e4e9f1;
-  border-radius: 8px;
-}
-
-.record-table {
-  width: 100%;
-}
-
-.record-table :deep(.el-table__header th) {
-  height: 44px;
-  background: #f5f7fa;
-  color: #475467;
-  font-size: 12px;
-  font-weight: 600;
-}
-
-.record-table :deep(.el-table__cell) {
-  padding: 10px 0;
-}
-
-.record-table :deep(.record-table-row-invalid td) {
-  background: #fff7f7 !important;
-}
-
-.table-index {
-  display: flex;
-  align-items: center;
-  flex-direction: column;
-  gap: 1px;
-  color: #667085;
-  font-variant-numeric: tabular-nums;
-}
-
-.table-index strong {
-  font-size: 13px;
-  font-weight: 600;
-}
-
-.table-index small {
-  color: #98a2b3;
-  font-size: 10px;
-}
-
-.table-cell-text {
-  display: -webkit-box;
-  overflow: hidden;
-  color: #526076;
-  font-size: 12px;
-  line-height: 1.6;
-  overflow-wrap: anywhere;
-  -webkit-box-orient: vertical;
-  -webkit-line-clamp: 2;
-}
-
-.record-table .el-button i {
-  margin-right: 5px;
-}
-
-.records-pagination {
-  display: flex;
-  justify-content: flex-end;
-  margin-top: 16px;
-}
-
-.record-empty-state {
-  min-height: 320px;
-}
-
-.code-viewer {
-  height: 100%;
-  max-height: 468px;
-  overflow: auto;
-  padding: 12px 0 18px;
-  outline: none;
-}
-
-.code-viewer:focus-visible {
-  box-shadow: inset 0 0 0 2px var(--el-color-primary-light-5);
-}
-
-.code-line {
-  display: grid;
-  grid-template-columns: 54px minmax(0, 1fr);
-  min-height: 25px;
-  color: #344054;
-  font-family: 'SFMono-Regular', Consolas, 'Liberation Mono', monospace;
-  font-size: 12px;
-  line-height: 1.65;
-}
-
-.code-line:hover {
-  background: #f4f6fa;
-}
-
-.line-number {
-  padding-right: 14px;
-  border-right: 1px solid #edf0f5;
-  color: #a4adbc;
-  font-variant-numeric: tabular-nums;
-  text-align: right;
-  user-select: none;
-}
-
-.code-line code {
-  display: block;
-  min-width: max-content;
-  padding: 0 18px;
-  white-space: pre;
-}
-
 .viewer-state {
   display: flex;
   align-items: center;
@@ -1057,47 +686,6 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
   margin-right: 5px;
 }
 
-.editor-hint {
-  margin-bottom: 18px;
-  padding: 10px 12px;
-  border: 1px solid #dbe2ff;
-  border-radius: 8px;
-  background: #f5f7ff;
-  color: #56627a;
-  font-size: 13px;
-}
-
-.editor-hint i {
-  margin-right: 6px;
-  color: var(--primary-color);
-}
-
-.field-helper {
-  margin-top: 5px;
-  color: #8a94a6;
-  line-height: 1.5;
-}
-
-.raw-record-editor label {
-  display: block;
-  margin-bottom: 8px;
-  color: #344054;
-  font-size: 13px;
-  font-weight: 600;
-}
-
-.record-editor-actions {
-  display: flex;
-  align-items: center;
-  justify-content: space-between;
-  gap: 16px;
-}
-
-.record-editor-actions > span {
-  color: #8a94a6;
-  font-size: 12px;
-}
-
 .empty-files {
   display: flex;
   align-items: center;
@@ -1149,18 +737,6 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
     grid-template-columns: 1fr;
   }
 
-  .version-control-bar,
-  .version-selector-group,
-  .version-status {
-    align-items: stretch;
-    flex-direction: column;
-  }
-
-  .version-select,
-  .version-control-bar > .el-button {
-    width: 100%;
-  }
-
   .overview-item {
     padding: 0 0 12px;
     border-right: 0;
@@ -1176,22 +752,6 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
     align-items: flex-start;
   }
 
-  .records-toolbar,
-  .record-editor-actions {
-    align-items: stretch;
-    flex-direction: column;
-  }
-
-  .record-search {
-    width: 100%;
-  }
-
-  .record-editor-actions > div {
-    display: grid;
-    grid-template-columns: 1fr 1fr;
-    gap: 8px;
-  }
-
   .download-button {
     width: 44px;
     padding: 8px;
@@ -1204,7 +764,4 @@ onBeforeUnmount(() => window.removeEventListener('beforeunload', handleBeforeUnl
   }
 }
 
-@media (prefers-reduced-motion: reduce) {
-  .record-table :deep(.el-table__row) { transition: none; }
-}
 </style>

frontend/src/views/dataset/preview/DatasetRawPreview.vue

@@ -0,0 +1,63 @@
+<script setup lang="ts">
+import { computed } from 'vue'
+
+const props = defineProps<{
+  content: string
+  fileName?: string
+}>()
+
+const lines = computed(() => props.content.split('\n').slice(0, 100))
+</script>
+
+<template>
+  <div class="code-viewer" tabindex="0" :aria-label="`${fileName || ''} 文件内容`">
+    <div v-for="(line, index) in lines" :key="index" class="code-line">
+      <span class="line-number" aria-hidden="true">{{ index + 1 }}</span>
+      <code>{{ line || ' ' }}</code>
+    </div>
+  </div>
+</template>
+
+<style scoped lang="scss">
+.code-viewer {
+  height: 100%;
+  max-height: 468px;
+  overflow: auto;
+  padding: 12px 0 18px;
+  outline: none;
+}
+
+.code-viewer:focus-visible {
+  box-shadow: inset 0 0 0 2px var(--el-color-primary-light-5);
+}
+
+.code-line {
+  display: grid;
+  grid-template-columns: 54px minmax(0, 1fr);
+  min-height: 25px;
+  color: #344054;
+  font-family: 'SFMono-Regular', Consolas, 'Liberation Mono', monospace;
+  font-size: 12px;
+  line-height: 1.65;
+}
+
+.code-line:hover {
+  background: #f4f6fa;
+}
+
+.line-number {
+  padding-right: 14px;
+  border-right: 1px solid #edf0f5;
+  color: #a4adbc;
+  font-variant-numeric: tabular-nums;
+  text-align: right;
+  user-select: none;
+}
+
+.code-line code {
+  display: block;
+  min-width: max-content;
+  padding: 0 18px;
+  white-space: pre;
+}
+</style>

frontend/src/views/dataset/preview/DatasetRecordEditorDialog.vue

@@ -0,0 +1,155 @@
+<script setup lang="ts">
+import type { DatasetRecord } from '../datasetRecords'
+import type { EditField } from './types'
+
+defineProps<{
+  modelValue: boolean
+  record: DatasetRecord | null
+  editFields: EditField[]
+  rawDraft: string
+  savingRecord: boolean
+  hasDraftChanges: boolean
+  nextVersionNumber: number
+}>()
+
+const emit = defineEmits<{
+  close: [done?: () => void]
+  save: []
+  'update:field': [index: number, value: string]
+  'update:rawDraft': [value: string]
+}>()
+
+function handleBeforeClose(done: () => void) {
+  emit('close', done)
+}
+
+function handleKeydown(event: KeyboardEvent) {
+  if ((event.ctrlKey || event.metaKey) && event.key.toLowerCase() === 's') {
+    event.preventDefault()
+    emit('save')
+  }
+}
+</script>
+
+<template>
+  <el-dialog
+    :model-value="modelValue"
+    class="record-editor-dialog"
+    width="min(720px, 92vw)"
+    :title="record ? `编辑第 ${record.displayIndex} 条数据` : '编辑数据'"
+    :close-on-click-modal="false"
+    :close-on-press-escape="!savingRecord"
+    :show-close="!savingRecord"
+    :before-close="handleBeforeClose"
+    destroy-on-close
+  >
+    <div v-if="record" class="record-editor" @keydown="handleKeydown">
+      <div class="editor-hint">
+        <i class="fa fa-info-circle" aria-hidden="true" />
+        当前修改会先暂存；完成多条修改后，点击文件工具栏中的“保存版本”创建
+        V{{ nextVersionNumber }}。
+      </div>
+      <el-form v-if="record.kind === 'object'" label-position="top">
+        <el-form-item v-for="(field, index) in editFields" :key="field.key" :label="field.key">
+          <el-input
+            :model-value="field.value"
+            type="textarea"
+            :rows="field.isJson ? 5 : 3"
+            :aria-label="`编辑字段 ${field.key}`"
+            :placeholder="field.isJson ? '请输入合法 JSON' : `请输入 ${field.key}`"
+            @update:model-value="emit('update:field', index, $event)"
+          />
+          <small v-if="field.isJson" class="field-helper">
+            该字段为对象、数组或数值，请保持合法 JSON 格式。
+          </small>
+        </el-form-item>
+      </el-form>
+      <div v-else class="raw-record-editor">
+        <label for="raw-record-draft">
+          {{ record.kind === 'invalid' ? '修复 JSON 内容' : 'JSON 内容' }}
+        </label>
+        <el-input
+          id="raw-record-draft"
+          :model-value="rawDraft"
+          type="textarea"
+          :rows="10"
+          aria-label="编辑当前条 JSON 内容"
+          placeholder="请输入合法 JSON"
+          @update:model-value="emit('update:rawDraft', $event)"
+        />
+      </div>
+    </div>
+    <template #footer>
+      <div class="record-editor-actions">
+        <span>{{ hasDraftChanges ? '有未暂存修改' : '尚未修改' }} · Ctrl/⌘ + S 暂存</span>
+        <div>
+          <el-button :disabled="savingRecord" @click="emit('close')">取消</el-button>
+          <el-button
+            type="primary"
+            :loading="savingRecord"
+            :disabled="!hasDraftChanges"
+            @click="emit('save')"
+          >
+            暂存修改
+          </el-button>
+        </div>
+      </div>
+    </template>
+  </el-dialog>
+</template>
+
+<style scoped lang="scss">
+.editor-hint {
+  margin-bottom: 18px;
+  padding: 10px 12px;
+  border: 1px solid #dbe2ff;
+  border-radius: 8px;
+  background: #f5f7ff;
+  color: #56627a;
+  font-size: 13px;
+}
+
+.editor-hint i {
+  margin-right: 6px;
+  color: var(--primary-color);
+}
+
+.field-helper {
+  margin-top: 5px;
+  color: #8a94a6;
+  line-height: 1.5;
+}
+
+.raw-record-editor label {
+  display: block;
+  margin-bottom: 8px;
+  color: #344054;
+  font-size: 13px;
+  font-weight: 600;
+}
+
+.record-editor-actions {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 16px;
+}
+
+.record-editor-actions > span {
+  color: #8a94a6;
+  font-size: 12px;
+}
+
+@media (max-width: 680px) {
+  .record-editor-actions {
+    align-items: stretch;
+    flex-direction: column;
+  }
+
+  .record-editor-actions > div {
+    display: grid;
+    grid-template-columns: 1fr 1fr;
+    gap: 8px;
+  }
+}
+</style>

frontend/src/views/dataset/preview/DatasetRecordTable.vue

@@ -0,0 +1,335 @@
+<script setup lang="ts">
+import { computed, ref, watch } from 'vue'
+import type { DatasetRecord } from '../datasetRecords'
+
+const props = defineProps<{
+  records: DatasetRecord[]
+  isViewingActiveVersion: boolean
+  versionId: string
+}>()
+
+const emit = defineEmits<{
+  edit: [record: DatasetRecord]
+}>()
+
+const pageSize = 10
+const searchText = ref('')
+const currentPage = ref(1)
+
+const allTableFieldKeys = computed(() => {
+  const keys = new Set<string>()
+  props.records.forEach((record) => {
+    if (record.kind !== 'object') return
+    Object.keys(record.value as Record<string, unknown>).forEach((key) => keys.add(key))
+  })
+  const preferred = ['instruction', 'question', 'prompt', 'input', 'context', 'output', 'answer', 'response']
+  return [...keys].sort((a, b) => {
+    const rank = (key: string) => {
+      const index = preferred.indexOf(key)
+      return index === -1 ? preferred.length : index
+    }
+    return rank(a) - rank(b)
+  })
+})
+const tableFieldKeys = computed(() => allTableFieldKeys.value.slice(0, 6))
+const invalidRecordCount = computed(() => props.records.filter((record) => record.kind === 'invalid').length)
+const hasExtraTableFields = computed(() => allTableFieldKeys.value.length > tableFieldKeys.value.length)
+const filteredRecords = computed(() => {
+  const query = searchText.value.trim().toLowerCase()
+  if (!query) return props.records
+  return props.records.filter((record) => JSON.stringify(record.value).toLowerCase().includes(query))
+})
+const pagedRecords = computed(() => {
+  const start = (currentPage.value - 1) * pageSize
+  return filteredRecords.value.slice(start, start + pageSize)
+})
+
+watch(
+  () => props.versionId,
+  () => {
+    searchText.value = ''
+    currentPage.value = 1
+  },
+)
+
+function asDatasetRecord(row: unknown) {
+  return row as DatasetRecord
+}
+
+function formatFieldValue(value: unknown) {
+  if (typeof value === 'string') return value || '（空）'
+  return JSON.stringify(value)
+}
+
+function recordFieldValue(record: DatasetRecord, key: string) {
+  if (record.kind !== 'object') {
+    return record.kind === 'invalid' ? record.raw : formatFieldValue(record.value)
+  }
+  const data = record.value as Record<string, unknown>
+  return Object.prototype.hasOwnProperty.call(data, key) ? formatFieldValue(data[key]) : '-'
+}
+
+function extraFieldValue(record: DatasetRecord) {
+  if (record.kind !== 'object') return '-'
+  const data = record.value as Record<string, unknown>
+  return allTableFieldKeys.value
+    .filter((key) => !tableFieldKeys.value.includes(key) && Object.prototype.hasOwnProperty.call(data, key))
+    .map((key) => `${key}: ${formatFieldValue(data[key])}`)
+    .join('；') || '-'
+}
+
+function recordRowClassName({ row }: { row: unknown }) {
+  return asDatasetRecord(row).kind === 'invalid' ? 'record-table-row-invalid' : ''
+}
+</script>
+
+<template>
+  <div class="records-viewer">
+    <div class="records-toolbar">
+      <div>
+        <strong>样本数据</strong>
+        <span>共 {{ records.length.toLocaleString() }} 条</span>
+        <el-tag v-if="invalidRecordCount" size="small" type="danger">
+          {{ invalidRecordCount }} 条格式错误
+        </el-tag>
+      </div>
+      <el-input
+        v-model="searchText"
+        class="record-search"
+        clearable
+        placeholder="搜索样本内容"
+        aria-label="搜索样本内容"
+        @input="currentPage = 1"
+      >
+        <template #prefix><i class="fa fa-search" /></template>
+      </el-input>
+    </div>
+
+    <div v-if="!filteredRecords.length" class="viewer-state record-empty-state">
+      <i class="fa fa-search" aria-hidden="true" />
+      <span>没有找到匹配的样本</span>
+    </div>
+    <div v-else class="record-table-shell" aria-label="数据样本表格">
+      <el-table
+        :data="pagedRecords"
+        row-key="sourceIndex"
+        stripe
+        border
+        class="record-table"
+        :row-class-name="recordRowClassName"
+      >
+        <el-table-column label="序号" width="76" align="center">
+          <template #default="{ row }">
+            <div class="table-index">
+              <strong>{{ asDatasetRecord(row).displayIndex }}</strong>
+              <small>第 {{ asDatasetRecord(row).sourceIndex + 1 }} 行</small>
+            </div>
+          </template>
+        </el-table-column>
+        <el-table-column
+          v-for="fieldKey in tableFieldKeys"
+          :key="fieldKey"
+          :label="fieldKey"
+          min-width="220"
+        >
+          <template #default="{ row }">
+            <el-tooltip
+              :content="recordFieldValue(asDatasetRecord(row), fieldKey)"
+              placement="top"
+              :show-after="400"
+            >
+              <div class="table-cell-text">
+                {{ recordFieldValue(asDatasetRecord(row), fieldKey) }}
+              </div>
+            </el-tooltip>
+          </template>
+        </el-table-column>
+        <el-table-column v-if="hasExtraTableFields" label="其他字段" min-width="240">
+          <template #default="{ row }">
+            <el-tooltip
+              :content="extraFieldValue(asDatasetRecord(row))"
+              placement="top"
+              :show-after="400"
+            >
+              <div class="table-cell-text">{{ extraFieldValue(asDatasetRecord(row)) }}</div>
+            </el-tooltip>
+          </template>
+        </el-table-column>
+        <el-table-column label="状态" width="96" align="center">
+          <template #default="{ row }">
+            <el-tag v-if="asDatasetRecord(row).kind === 'invalid'" size="small" type="danger">
+              格式错误
+            </el-tag>
+            <el-tag v-else-if="asDatasetRecord(row).kind === 'primitive'" size="small" type="warning">
+              非对象
+            </el-tag>
+            <el-tag v-else size="small" type="success">正常</el-tag>
+          </template>
+        </el-table-column>
+        <el-table-column label="操作" width="96" align="center" fixed="right">
+          <template #default="{ row }">
+            <el-button
+              type="primary"
+              link
+              :disabled="!isViewingActiveVersion"
+              :title="isViewingActiveVersion ? '' : '历史版本为只读，请先设为当前版本'"
+              :aria-label="`${asDatasetRecord(row).kind === 'invalid' ? '修复' : '编辑'}第 ${asDatasetRecord(row).displayIndex} 条数据`"
+              @click="emit('edit', asDatasetRecord(row))"
+            >
+              <i :class="asDatasetRecord(row).kind === 'invalid' ? 'fa fa-wrench' : 'fa fa-pencil'" />
+              {{ asDatasetRecord(row).kind === 'invalid' ? '修复' : '编辑' }}
+            </el-button>
+          </template>
+        </el-table-column>
+      </el-table>
+    </div>
+    <div v-if="filteredRecords.length > pageSize" class="records-pagination">
+      <el-pagination
+        v-model:current-page="currentPage"
+        :page-size="pageSize"
+        :total="filteredRecords.length"
+        layout="prev, pager, next"
+        background
+      />
+    </div>
+  </div>
+</template>
+
+<style scoped lang="scss">
+.records-viewer {
+  min-height: 468px;
+  padding: 16px;
+  box-sizing: border-box;
+}
+
+.records-toolbar {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 16px;
+  margin-bottom: 14px;
+}
+
+.records-toolbar > div {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  min-width: 0;
+}
+
+.records-toolbar strong {
+  color: #253047;
+  font-size: 14px;
+}
+
+.records-toolbar span {
+  color: #8a94a6;
+  font-size: 12px;
+}
+
+.record-search {
+  width: 240px;
+}
+
+.record-table-shell {
+  overflow: hidden;
+  border: 1px solid #e4e9f1;
+  border-radius: 8px;
+}
+
+.record-table {
+  width: 100%;
+}
+
+.record-table :deep(.el-table__header th) {
+  height: 44px;
+  background: #f5f7fa;
+  color: #475467;
+  font-size: 12px;
+  font-weight: 600;
+}
+
+.record-table :deep(.el-table__cell) {
+  padding: 10px 0;
+}
+
+.record-table :deep(.record-table-row-invalid td) {
+  background: #fff7f7 !important;
+}
+
+.table-index {
+  display: flex;
+  align-items: center;
+  flex-direction: column;
+  gap: 1px;
+  color: #667085;
+  font-variant-numeric: tabular-nums;
+}
+
+.table-index strong {
+  font-size: 13px;
+  font-weight: 600;
+}
+
+.table-index small {
+  color: #98a2b3;
+  font-size: 10px;
+}
+
+.table-cell-text {
+  display: -webkit-box;
+  overflow: hidden;
+  color: #526076;
+  font-size: 12px;
+  line-height: 1.6;
+  overflow-wrap: anywhere;
+  -webkit-box-orient: vertical;
+  -webkit-line-clamp: 2;
+}
+
+.record-table .el-button i {
+  margin-right: 5px;
+}
+
+.records-pagination {
+  display: flex;
+  justify-content: flex-end;
+  margin-top: 16px;
+}
+
+.viewer-state {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  flex-direction: column;
+  gap: 10px;
+  height: 100%;
+  min-height: 400px;
+  color: #8b96a8;
+  font-size: 13px;
+}
+
+.viewer-state > i {
+  color: #aeb7c5;
+  font-size: 24px;
+}
+
+.record-empty-state {
+  min-height: 320px;
+}
+
+@media (max-width: 680px) {
+  .records-toolbar {
+    align-items: stretch;
+    flex-direction: column;
+  }
+
+  .record-search {
+    width: 100%;
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .record-table :deep(.el-table__row) { transition: none; }
+}
+</style>

frontend/src/views/dataset/preview/DatasetVersionBar.vue

@@ -0,0 +1,154 @@
+<script setup lang="ts">
+import type { DatasetVersion } from '@/types'
+
+const props = defineProps<{
+  modelValue: string
+  versions: DatasetVersion[]
+  viewedVersion?: DatasetVersion
+  activeVersionId: string
+  isViewingActiveVersion: boolean
+  hasPendingVersionChanges: boolean
+  versionLoading: boolean
+  savingRecord: boolean
+  activatingVersion: boolean
+}>()
+
+const emit = defineEmits<{
+  'update:modelValue': [value: string]
+  change: [versionId: string]
+  activate: []
+}>()
+
+function formatVersionTime(value: string) {
+  return new Date(value).toLocaleString('zh-CN', { hour12: false })
+}
+
+function versionOptionLabel(version: DatasetVersion) {
+  const status = version.id === props.activeVersionId ? '当前版本' : '历史版本'
+  return `V${version.version} · ${status} · ${formatVersionTime(version.create_time)}`
+}
+
+function handleVersionChange(versionId: string) {
+  emit('update:modelValue', versionId)
+  emit('change', versionId)
+}
+</script>
+
+<template>
+  <section class="version-control-bar" aria-label="数据集版本控制">
+    <div class="version-selector-group">
+      <label for="dataset-version-select">数据版本</label>
+      <el-select
+        id="dataset-version-select"
+        :model-value="modelValue"
+        class="version-select"
+        :loading="versionLoading"
+        :disabled="versionLoading || savingRecord"
+        aria-label="选择要查看的数据集版本"
+        @change="handleVersionChange"
+      >
+        <el-option
+          v-for="version in versions"
+          :key="version.id"
+          :label="versionOptionLabel(version)"
+          :value="version.id"
+        />
+      </el-select>
+    </div>
+
+    <div class="version-status" role="status" aria-live="polite">
+      <template v-if="viewedVersion">
+        <el-tag v-if="isViewingActiveVersion" type="success">
+          V{{ viewedVersion.version }} · 当前版本
+        </el-tag>
+        <el-tag v-else type="info">V{{ viewedVersion.version }} · 历史版本（只读）</el-tag>
+        <el-tag v-if="hasPendingVersionChanges" type="warning">有待保存修改</el-tag>
+        <span>
+          {{ viewedVersion.description || '无版本说明' }} ·
+          {{ formatVersionTime(viewedVersion.create_time) }}
+        </span>
+      </template>
+    </div>
+
+    <el-button
+      v-if="viewedVersion && !isViewingActiveVersion"
+      type="primary"
+      :loading="activatingVersion"
+      @click="emit('activate')"
+    >
+      <i class="fa fa-check-circle" />
+      设为当前版本
+    </el-button>
+  </section>
+</template>
+
+<style scoped lang="scss">
+.version-control-bar {
+  display: flex;
+  align-items: center;
+  gap: 16px;
+  min-height: 64px;
+  margin-bottom: 12px;
+  padding: 10px 14px;
+  box-sizing: border-box;
+  border: 1px solid #dfe5ee;
+  border-radius: 10px;
+  background: #fff;
+}
+
+.version-selector-group {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  flex: 0 0 auto;
+}
+
+.version-selector-group label {
+  color: #344054;
+  font-size: 13px;
+  font-weight: 600;
+}
+
+.version-select {
+  width: 300px;
+}
+
+.version-status {
+  display: flex;
+  align-items: center;
+  flex: 1;
+  gap: 10px;
+  min-width: 0;
+}
+
+.version-status > span {
+  overflow: hidden;
+  color: #7c8799;
+  font-size: 12px;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.version-control-bar > .el-button {
+  min-height: 40px;
+  flex: 0 0 auto;
+}
+
+.version-control-bar > .el-button i {
+  margin-right: 6px;
+}
+
+@media (max-width: 680px) {
+  .version-control-bar,
+  .version-selector-group,
+  .version-status {
+    align-items: stretch;
+    flex-direction: column;
+  }
+
+  .version-select,
+  .version-control-bar > .el-button {
+    width: 100%;
+  }
+}
+</style>

frontend/src/views/dataset/preview/types.ts

@@ -0,0 +1,5 @@
+export interface EditField {
+  key: string
+  value: string
+  isJson: boolean
+}

frontend/src/views/fine-tune/FineTuneCreateView.vue

@@ -14,6 +14,14 @@ import { getModelList } from '@/api/modules/model'
 import { getDatasetList } from '@/api/modules/dataset'
 import { getSystemInfo } from '@/api/modules/system'
 import { TEMPLATE_GROUPS, LR_SCHEDULER_OPTIONS, QUANTIZATION_BIT_OPTIONS, QUANT_METHOD_OPTIONS, GGUF_FORMAT_OPTIONS } from '@/constants'
+import {
+  DEFAULT_TRAINING_PARAMS,
+  buildFineTuneCommand,
+  buildFineTunePayload,
+  createDefaultFineTuneForm,
+  toCreateFineTunePayload,
+} from './fineTuneFormModel'
+import type { FineTuneFormModel } from './fineTuneFormModel'
 import type { ModelItem, DatasetItem, GpuInfo } from '@/types'
 
 const router = useRouter()
@@ -26,36 +34,7 @@ const gpus = ref<GpuInfo[]>([])
 const selectedGpus = ref<number[]>([])
 const modelDialogVisible = ref(false)
 
-const form = reactive({
+const form = reactive(createDefaultFineTuneForm())
-  name: '',
-  description: '',
-  train_type: 'SFT' as 'SFT' | 'DPO' | 'CPT',
-  base_model: '' as string | number,
-  template: 'qwen',
-  train_method: 'lora' as 'lora' | 'full',
-  train_dataset_id: '' as string | number,
-  auto_merge: false,
-  // 训练参数
-  batch_size: 1,
-  learning_rate: 0.0001,
-  n_epochs: 1,
-  save_steps: 100,
-  lr_scheduler_type: 'cosine',
-  max_length: 512,
-  warmup_ratio: 0.05,
-  weight_decay: 0.01,
-  // LoRA 参数
-  lora_alpha: 16, // 修复原项目 lora_alpha 默认值不一致 bug
-  lora_dropout: 0.1,
-  lora_rank: 8,
-  // 量化参数
-  quantization_bit: 0, // 训练时量化（QLoRA）：0=不量化
-  export_quantized: false, // 训练后是否导出量化模型
-  quant_method: 'bnb', // 导出量化方法
-  quant_bits: 4, // 导出量化位数
-  quant_group_size: 128, // 分组大小（GPTQ/AWQ）
-  export_format: 'Q4_K_M', // GGUF 导出格式
-})
 
 const rules: FormRules = {
   name: [
@@ -78,45 +57,8 @@ const selectedModel = computed(() => models.value.find((model) => model.id === f
 
 const modelDialogTitle = computed(() => selectedModel.value?.name || '')
 
-/** 训练命令实时预览 */
+/** 训练命令与提交载荷共用同一份表单模型。 */
-const commandPreview = computed(() => {
+const commandPreview = computed(() => buildFineTuneCommand(form, selectedGpus.value))
-  const gpuIds = selectedGpus.value.length ? selectedGpus.value.join(',') : '0'
-  let cmd = `CUDA_VISIBLE_DEVICES=${gpuIds} llamafactory-cli train \\\n`
-  cmd += `  --stage ${form.train_type === 'DPO' ? 'dpo' : form.train_type === 'CPT' ? 'cpt' : 'sft'} \\\n`
-  cmd += `  --do_train \\\n`
-  cmd += `  --model_name_or_path <base_model_path> \\\n`
-  cmd += `  --dataset <dataset> \\\n`
-  cmd += `  --template ${form.template} \\\n`
-  cmd += `  --finetuning_type ${form.train_method} \\\n`
-  cmd += `  --output_dir ./saves/${form.name || 'output'} \\\n`
-  cmd += `  --per_device_train_batch_size ${form.batch_size} \\\n`
-  cmd += `  --learning_rate ${form.learning_rate} \\\n`
-  cmd += `  --num_train_epochs ${form.n_epochs} \\\n`
-  cmd += `  --save_steps ${form.save_steps} \\\n`
-  cmd += `  --lr_scheduler_type ${form.lr_scheduler_type} \\\n`
-  cmd += `  --cutoff_len ${form.max_length} \\\n`
-  cmd += `  --warmup_ratio ${form.warmup_ratio} \\\n`
-  cmd += `  --weight_decay ${form.weight_decay}`
-  if (showLoraParams.value) {
-    cmd += ` \\\n  --lora_alpha ${form.lora_alpha}`
-    cmd += ` \\\n  --lora_dropout ${form.lora_dropout}`
-    cmd += ` \\\n  --lora_rank ${form.lora_rank}`
-  }
-  if (showLoraParams.value && form.quantization_bit) {
-    cmd += ` \\\n  --quantization_bit ${form.quantization_bit}`
-  }
-  if (form.export_quantized) {
-    const method = form.quant_method
-    const bits = form.quant_bits
-    cmd += ` \\\n  # 训练后导出量化模型：${method} ${bits}bit`
-    if (method === 'gguf') {
-      cmd += ` \\\n  # export_format=${form.export_format}`
-    } else if (method === 'gptq' || method === 'awq') {
-      cmd += ` \\\n  # group_size=${form.quant_group_size}`
-    }
-  }
-  return cmd
-})
 
 /** GPU 多选切换 */
 function toggleGpu(index: number) {
@@ -140,31 +82,26 @@ function handleModelConfirm(modelId: string | number) {
 }
 
 function resetParams() {
-  Object.assign(form, {
+  Object.assign(form, DEFAULT_TRAINING_PARAMS)
-    batch_size: 1,
-    learning_rate: 0.0001,
-    n_epochs: 1,
-    save_steps: 100,
-    lr_scheduler_type: 'cosine',
-    max_length: 512,
-    warmup_ratio: 0.05,
-    weight_decay: 0.01,
-    lora_alpha: 16,
-    lora_dropout: 0.1,
-    lora_rank: 8,
-    quantization_bit: 0,
-    export_quantized: false,
-    quant_method: 'bnb',
-    quant_bits: 4,
-    quant_group_size: 128,
-    export_format: 'Q4_K_M',
-  })
 }
 
 const isParamsExpanded = ref(false)
 
+interface ParameterDefinition {
+  key: keyof FineTuneFormModel
+  name: string
+  desc: string
+  hint: string
+  type: 'number' | 'select'
+  min?: number
+  max?: number
+  step?: number
+  precision?: number
+  options?: Array<{ label: string; value: string | number }>
+}
+
 const allParams = computed(() => {
-  const params = [
+  const params: ParameterDefinition[] = [
     { key: 'batch_size', name: 'batch_size', desc: '批次大小，代表模型训练过程中，模型更新一次参数所需要的数据样本数。', hint: '[1, 64], step:1', type: 'number', min: 1, max: 64, step: 1 },
     { key: 'learning_rate', name: 'learning_rate', desc: '学习率，代表每次更新数据的增量参数权重比例。', hint: '[0.000001, 1]', type: 'number', min: 0.000001, max: 1, step: 0.00001, precision: 6 },
     { key: 'n_epochs', name: 'n_epochs', desc: '循环次数，代表模型训练过程中模型学习数据集的次数，可理解为看几遍数据，一般建议的范围是 1-3 遍即可，可依据需求进行调整', hint: '[1, 100], step:1', type: 'number', min: 1, max: 100, step: 1 },
@@ -178,7 +115,7 @@ const allParams = computed(() => {
     params.push(
       { key: 'lora_alpha', name: 'lora_alpha', desc: 'LoRA 缩放系数。', hint: '16/32/64/128', type: 'select', options: [{ label: '16', value: 16 }, { label: '32', value: 32 }, { label: '64', value: 64 }, { label: '128', value: 128 }] },
       { key: 'lora_rank', name: 'lora_rank', desc: 'LoRA 秩大小，控制低秩矩阵的维度。', hint: '8/16/32/64', type: 'select', options: [{ label: '8', value: 8 }, { label: '16', value: 16 }, { label: '32', value: 32 }, { label: '64', value: 64 }] },
-      { key: 'lora_dropout', name: 'lora_dropout', desc: 'LoRA 层的 dropout 比例。', hint: '[0, 1]', type: 'number', min: 0, max: 1, step: 0.05, precision: 2 }
+      { key: 'lora_dropout', name: 'lora_dropout', desc: 'LoRA 层的 dropout 比例。', hint: '[0, 1]', type: 'number', min: 0, max: 1, step: 0.05, precision: 2 },
     )
   }
   return params
@@ -188,6 +125,15 @@ const visibleParams = computed(() => {
   return isParamsExpanded.value ? allParams.value : allParams.value.slice(0, 3)
 })
 
+function parameterValue(key: keyof FineTuneFormModel) {
+  const value = form[key]
+  return typeof value === 'boolean' ? Number(value) : value
+}
+
+function updateParameterValue(key: keyof FineTuneFormModel, value: string | number | undefined) {
+  if (value !== undefined) Reflect.set(form, key, value)
+}
+
 async function loadModels() {
   try {
     models.value = (await getModelList()) || []
@@ -225,88 +171,32 @@ async function handleSubmit() {
     }
     submitting.value = true
     try {
-      // 任务名查重
+      let check: { exists: boolean }
-      const check = await checkFineTuneName(form.name).catch(() => ({ exists: false }))
+      try {
-      if ((check as any).exists) {
+        check = await checkFineTuneName(form.name)
+      } catch {
+        ElMessage.error('任务名校验失败，请稍后重试')
+        return
+      }
+      if (check.exists) {
         ElMessage.error('任务名称已存在，请更换')
-        submitting.value = false
         return
       }
 
-      // 第一步：创建任务记录
+      const payload = buildFineTunePayload(form, selectedGpus.value)
-      const taskData = {
+      const createRes = await createFineTune(toCreateFineTunePayload(payload))
-        name: form.name,
+      const taskId = createRes.id
-        description: form.description,
-        base_model: form.base_model,
-        template: form.template,
-        train_type: form.train_type,
-        train_method: form.train_method,
-        gpus: selectedGpus.value,
-        train_dataset_id: form.train_dataset_id,
-        auto_merge: form.train_type === 'SFT' && form.auto_merge,
-        output_model_name: form.name,
-        batch_size: form.batch_size,
-        learning_rate: form.learning_rate,
-        n_epochs: form.n_epochs,
-        save_steps: form.save_steps,
-        lr_scheduler_type: form.lr_scheduler_type,
-        max_length: form.max_length,
-        warmup_ratio: form.warmup_ratio,
-        weight_decay: form.weight_decay,
-        lora_alpha: form.lora_alpha,
-        lora_dropout: form.lora_dropout,
-        lora_rank: form.lora_rank,
-        quantization_bit: form.train_method === 'lora' ? form.quantization_bit : 0,
-        export_quantized: form.export_quantized,
-        quant_method: form.export_quantized ? form.quant_method : '',
-        quant_bits: form.export_quantized ? form.quant_bits : 0,
-        quant_group_size: form.export_quantized ? form.quant_group_size : 0,
-        export_format: form.export_quantized && form.quant_method === 'gguf' ? form.export_format : '',
-        status: 'pending',
-        progress: 0,
-      }
-      const createRes: any = await createFineTune(taskData)
-      const taskId = createRes?.id || createRes
 
-      // 第二步：启动训练
       try {
-        await startFineTune({
+        await startFineTune({ ...payload, task_id: taskId })
-          task_id: taskId,
-          name: form.name,
-          base_model: form.base_model,
-          template: form.template,
-          train_type: form.train_type,
-          train_method: form.train_method,
-          train_dataset_id: form.train_dataset_id,
-          auto_merge: form.train_type === 'SFT' && form.auto_merge,
-          output_model_name: form.name,
-          gpus: selectedGpus.value,
-          batch_size: form.batch_size,
-          learning_rate: form.learning_rate,
-          n_epochs: form.n_epochs,
-          save_steps: form.save_steps,
-          lr_scheduler_type: form.lr_scheduler_type,
-          max_length: form.max_length,
-          warmup_ratio: form.warmup_ratio,
-          weight_decay: form.weight_decay,
-          lora_alpha: form.lora_alpha,
-          lora_dropout: form.lora_dropout,
-          lora_rank: form.lora_rank,
-          quantization_bit: form.train_method === 'lora' ? form.quantization_bit : 0,
-          export_quantized: form.export_quantized,
-          quant_method: form.export_quantized ? form.quant_method : '',
-          quant_bits: form.export_quantized ? form.quant_bits : 0,
-          quant_group_size: form.export_quantized ? form.quant_group_size : 0,
-          export_format: form.export_quantized && form.quant_method === 'gguf' ? form.export_format : '',
-        })
         ElMessage.success('训练任务已创建并启动')
-      } catch (e) {
+      } catch {
-        // 启动失败，回写状态
         await updateFineTune(taskId, { status: 'failed' })
+        ElMessage.error('任务已创建，但训练启动失败')
       }
       router.push('/fine-tune')
     } catch {
-      // ignore
+      ElMessage.error('训练任务创建失败，请稍后重试')
     } finally {
       submitting.value = false
     }
@@ -427,15 +317,17 @@ onMounted(() => {
                 <div class="param-col config">
                   <el-input-number 
                     v-if="param.type === 'number'" 
-                    v-model="(form as any)[param.key]" 
+                    :model-value="Number(parameterValue(param.key))"
                     :min="param.min" :max="param.max" :step="param.step" :precision="param.precision" 
                     controls-position="right"
                     style="width: 200px"
+                    @update:model-value="updateParameterValue(param.key, $event)"
                   />
                   <el-select 
                     v-else-if="param.type === 'select'" 
-                    v-model="(form as any)[param.key]" 
+                    :model-value="parameterValue(param.key)"
                     style="width: 200px"
+                    @update:model-value="updateParameterValue(param.key, $event)"
                   >
                     <el-option v-for="o in param.options" :key="o.value" :label="o.label" :value="o.value" />
                   </el-select>

frontend/src/views/fine-tune/FineTuneListView.vue

@@ -37,7 +37,7 @@ const filteredList = computed(() => {
     if (filters.value.trainType.length && !filters.value.trainType.includes(row.train_type)) {
       return false
     }
-    if (filters.value.trainMethod.length && !filters.value.trainMethod.includes(row.train_method)) {
+    if (filters.value.trainMethod.length && (!row.train_method || !filters.value.trainMethod.includes(row.train_method))) {
       return false
     }
     return true

frontend/src/views/fine-tune/fineTuneFormModel.ts

@@ -0,0 +1,146 @@
+import type { FineTuneStartPayload, FineTuneTask } from '@/types'
+
+export type FineTuneFormModel = {
+  name: string
+  description: string
+  train_type: 'SFT' | 'DPO' | 'CPT'
+  base_model: string | number
+  template: string
+  train_method: 'lora' | 'full'
+  train_dataset_id: string | number
+  auto_merge: boolean
+  batch_size: number
+  learning_rate: number
+  n_epochs: number
+  save_steps: number
+  lr_scheduler_type: string
+  max_length: number
+  warmup_ratio: number
+  weight_decay: number
+  lora_alpha: number
+  lora_dropout: number
+  lora_rank: number
+  quantization_bit: number
+  export_quantized: boolean
+  quant_method: string
+  quant_bits: number
+  quant_group_size: number
+  export_format: string
+}
+
+export const DEFAULT_TRAINING_PARAMS = {
+  batch_size: 1,
+  learning_rate: 0.0001,
+  n_epochs: 1,
+  save_steps: 100,
+  lr_scheduler_type: 'cosine',
+  max_length: 512,
+  warmup_ratio: 0.05,
+  weight_decay: 0.01,
+  lora_alpha: 16,
+  lora_dropout: 0.1,
+  lora_rank: 8,
+  quantization_bit: 0,
+  export_quantized: false,
+  quant_method: 'bnb',
+  quant_bits: 4,
+  quant_group_size: 128,
+  export_format: 'Q4_K_M',
+} as const
+
+export function createDefaultFineTuneForm(): FineTuneFormModel {
+  return {
+    name: '',
+    description: '',
+    train_type: 'SFT',
+    base_model: '',
+    template: 'qwen',
+    train_method: 'lora',
+    train_dataset_id: '',
+    auto_merge: false,
+    ...DEFAULT_TRAINING_PARAMS,
+  }
+}
+
+export function buildFineTunePayload(
+  form: FineTuneFormModel,
+  gpus: number[],
+): Omit<FineTuneStartPayload, 'task_id'> {
+  return {
+    name: form.name,
+    description: form.description,
+    base_model: form.base_model,
+    template: form.template,
+    train_type: form.train_type,
+    train_method: form.train_method,
+    gpus: [...gpus],
+    train_dataset_id: form.train_dataset_id,
+    auto_merge: form.train_type === 'SFT' && form.auto_merge,
+    output_model_name: form.name,
+    batch_size: form.batch_size,
+    learning_rate: form.learning_rate,
+    n_epochs: form.n_epochs,
+    save_steps: form.save_steps,
+    lr_scheduler_type: form.lr_scheduler_type,
+    max_length: form.max_length,
+    warmup_ratio: form.warmup_ratio,
+    weight_decay: form.weight_decay,
+    lora_alpha: form.lora_alpha,
+    lora_dropout: form.lora_dropout,
+    lora_rank: form.lora_rank,
+    quantization_bit: form.train_method === 'lora' ? form.quantization_bit : 0,
+    export_quantized: form.export_quantized,
+    quant_method: form.export_quantized ? form.quant_method : '',
+    quant_bits: form.export_quantized ? form.quant_bits : 0,
+    quant_group_size: form.export_quantized ? form.quant_group_size : 0,
+    export_format: form.export_quantized && form.quant_method === 'gguf' ? form.export_format : '',
+  }
+}
+
+export function buildFineTuneCommand(form: FineTuneFormModel, gpus: number[]) {
+  const gpuIds = gpus.length ? gpus.join(',') : '0'
+  const stage = form.train_type === 'DPO' ? 'dpo' : form.train_type === 'CPT' ? 'cpt' : 'sft'
+  const lines = [
+    `CUDA_VISIBLE_DEVICES=${gpuIds} llamafactory-cli train`,
+    `  --stage ${stage}`,
+    '  --do_train',
+    '  --model_name_or_path <base_model_path>',
+    '  --dataset <dataset>',
+    `  --template ${form.template}`,
+    `  --finetuning_type ${form.train_method}`,
+    `  --output_dir ./saves/${form.name || 'output'}`,
+    `  --per_device_train_batch_size ${form.batch_size}`,
+    `  --learning_rate ${form.learning_rate}`,
+    `  --num_train_epochs ${form.n_epochs}`,
+    `  --save_steps ${form.save_steps}`,
+    `  --lr_scheduler_type ${form.lr_scheduler_type}`,
+    `  --cutoff_len ${form.max_length}`,
+    `  --warmup_ratio ${form.warmup_ratio}`,
+    `  --weight_decay ${form.weight_decay}`,
+  ]
+
+  if (form.train_method === 'lora') {
+    lines.push(
+      `  --lora_alpha ${form.lora_alpha}`,
+      `  --lora_dropout ${form.lora_dropout}`,
+      `  --lora_rank ${form.lora_rank}`,
+    )
+    if (form.quantization_bit) lines.push(`  --quantization_bit ${form.quantization_bit}`)
+  }
+
+  if (form.export_quantized) {
+    lines.push(`  # 训练后导出量化模型：${form.quant_method} ${form.quant_bits}bit`)
+    if (form.quant_method === 'gguf') lines.push(`  # export_format=${form.export_format}`)
+    if (form.quant_method === 'gptq' || form.quant_method === 'awq') {
+      lines.push(`  # group_size=${form.quant_group_size}`)
+    }
+  }
+
+  return lines.join(' \\\n')
+}
+
+export function toCreateFineTunePayload(
+  payload: Omit<FineTuneStartPayload, 'task_id'>,
+): Partial<FineTuneTask> {
+  return { ...payload, status: 'pending', progress: 0 }
+}

frontend/src/views/system/training-log/TrainingTaskOverview.vue

@@ -0,0 +1,88 @@
+<script setup lang="ts">
+import PageCard from '@/components/PageCard.vue'
+import { DATASET_TYPE_MAP, STORAGE_MAP, TRAIN_METHOD_MAP, TRAIN_TYPE_MAP } from '@/constants'
+import type { DatasetItem, FineTuneTask } from '@/types'
+
+defineProps<{
+  task: FineTuneTask | null
+  dataset: DatasetItem | null
+  baseModelName: string
+}>()
+
+function formatDateTime(value?: string) {
+  if (!value) return '-'
+  const date = new Date(value)
+  if (Number.isNaN(date.getTime())) return value
+  return date.toLocaleString('zh-CN', { hour12: false })
+}
+</script>
+
+<template>
+  <PageCard class="task-overview" title="任务信息" subtitle="模型、数据集与运行配置">
+    <div class="overview-layout" aria-label="训练任务信息">
+      <el-descriptions class="task-descriptions task-profile" :column="2" border>
+        <el-descriptions-item label="任务 ID">{{ task?.id ?? '-' }}</el-descriptions-item>
+        <el-descriptions-item label="创建时间">{{ formatDateTime(task?.create_time) }}</el-descriptions-item>
+        <el-descriptions-item label="基座模型">{{ baseModelName }}</el-descriptions-item>
+        <el-descriptions-item label="输出模型">{{ task?.output_model_name || '暂未生成' }}</el-descriptions-item>
+        <el-descriptions-item label="训练方式">
+          {{ task?.train_type ? (TRAIN_TYPE_MAP[task.train_type] || task.train_type) : '未配置' }}
+        </el-descriptions-item>
+        <el-descriptions-item label="训练方法">
+          {{ task?.train_method ? (TRAIN_METHOD_MAP[task.train_method] || task.train_method) : '未配置' }}
+        </el-descriptions-item>
+        <el-descriptions-item label="训练数据集" class-name="dataset-profile">
+          {{ dataset?.name || (task?.train_dataset_id ? '正在加载' : '未配置') }}
+        </el-descriptions-item>
+        <el-descriptions-item label="数据类型">
+          {{ dataset ? (DATASET_TYPE_MAP[dataset.type] || dataset.type || '未配置') : '未配置' }}
+        </el-descriptions-item>
+        <el-descriptions-item label="数据条数">
+          {{ dataset?.count?.toLocaleString('zh-CN') ?? '未配置' }}{{ dataset?.count != null ? ' 条' : '' }}
+        </el-descriptions-item>
+        <el-descriptions-item label="数据大小">{{ dataset?.size || '未配置' }}</el-descriptions-item>
+        <el-descriptions-item label="训练开始时间" class-name="runtime-panel">
+          {{ formatDateTime(task?.create_time) }}
+        </el-descriptions-item>
+        <el-descriptions-item label="训练时长">{{ task?.train_duration || '未配置' }}</el-descriptions-item>
+        <el-descriptions-item label="存储位置">
+          {{ STORAGE_MAP[dataset?.storage_type || ''] || dataset?.storage_type || '未配置' }}
+        </el-descriptions-item>
+        <el-descriptions-item label="使用 GPU">
+          {{ task?.gpus?.length ? task.gpus.join('、') : '未配置' }}
+        </el-descriptions-item>
+      </el-descriptions>
+    </div>
+  </PageCard>
+</template>
+
+<style scoped lang="scss">
+.task-overview {
+  margin-bottom: 0;
+  border: 1px solid #e4e7ed !important;
+  border-radius: 8px !important;
+  box-shadow: none !important;
+}
+.overview-layout { width: 100%; }
+.task-descriptions :deep(.el-descriptions__label) {
+  width: 132px;
+  color: #606266;
+  font-weight: 500;
+  background: #f7f8fa !important;
+}
+.task-descriptions :deep(.el-descriptions__content) {
+  color: #303133;
+  font-weight: 500;
+  font-variant-numeric: tabular-nums;
+}
+.task-descriptions :deep(.el-descriptions__cell) { padding: 12px 16px !important; }
+
+@media (max-width: 700px) {
+  .task-descriptions :deep(.el-descriptions__body),
+  .task-descriptions :deep(.el-descriptions__table),
+  .task-descriptions :deep(.el-descriptions__tbody),
+  .task-descriptions :deep(.el-descriptions__row),
+  .task-descriptions :deep(.el-descriptions__cell) { display: block; width: 100%; box-sizing: border-box; }
+  .task-descriptions :deep(.el-descriptions__label) { width: 100%; border-bottom: 0 !important; }
+}
+</style>

frontend/src/views/system/training-log/trainingLogModel.ts

@@ -0,0 +1,165 @@
+import type { EChartsOption } from 'echarts'
+import type { FineTuneTask, TrainingLogFile } from '@/types'
+
+export interface TrainingMetricData {
+  loss: number[]
+  gradNorm: number[]
+  lr: number[]
+  epoch: number[]
+}
+
+export interface TrainingSummary {
+  epoch: string
+  trainLoss: string
+  runtime: string
+}
+
+export interface ParsedTrainingLog {
+  metrics: TrainingMetricData
+  summary: TrainingSummary
+}
+
+const NUMBER_SOURCE = '[-+]?(?:\\d+(?:\\.\\d*)?|\\.\\d+)(?:[eE][-+]?\\d+)?'
+
+function escapeRegExp(value: string) {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+}
+
+function extractNumber(source: string, key: string) {
+  const match = source.match(new RegExp(`['"]?${escapeRegExp(key)}['"]?\\s*:\\s*(${NUMBER_SOURCE})`, 'i'))
+  return match ? Number(match[1]) : undefined
+}
+
+function extractSummaryValue(source: string, key: string) {
+  const match = source.match(new RegExp(`['"]?${escapeRegExp(key)}['"]?\\s*(?:=|:)\\s*(${NUMBER_SOURCE})`, 'i'))
+  return match?.[1] || ''
+}
+
+/** 根据任务精确选择日志；PID 优先，任务名仅作为明确兜底。 */
+export function resolveTrainingLogFile(
+  files: TrainingLogFile[],
+  task: Pick<FineTuneTask, 'process_id' | 'name'>,
+) {
+  const processId = task.process_id
+  if (processId != null) {
+    const pidMatch = files.find((file) => file.pid === processId)
+    if (pidMatch) return pidMatch
+
+    const pidPattern = new RegExp(`(?:^|[^0-9])(?:pid)?${processId}(?:[^0-9]|$)`, 'i')
+    const filenameMatch = files.find((file) => pidPattern.test(file.file))
+    if (filenameMatch) return filenameMatch
+  }
+
+  const taskName = task.name.trim()
+  if (!taskName) return undefined
+  return files.find((file) => file.name.includes(taskName) || file.file.includes(taskName))
+}
+
+/** 解析日志中的逐步指标。字段顺序和常见数值格式均不受限制。 */
+export function parseTrainingMetrics(text: string): TrainingMetricData {
+  const metrics: TrainingMetricData = { loss: [], gradNorm: [], lr: [], epoch: [] }
+  const blocks = text.match(/\{[^{}\r\n]*\}/g) || []
+
+  for (const block of blocks) {
+    const loss = extractNumber(block, 'loss')
+    const gradNorm = extractNumber(block, 'grad_norm')
+    const learningRate = extractNumber(block, 'learning_rate')
+    const epoch = extractNumber(block, 'epoch')
+    if (loss == null || gradNorm == null || learningRate == null) continue
+    metrics.loss.push(loss)
+    metrics.gradNorm.push(gradNorm)
+    metrics.lr.push(learningRate)
+    if (epoch != null) metrics.epoch.push(epoch)
+  }
+
+  return metrics
+}
+
+/** 每次都返回新对象，日志截断或切换时不会残留上一轮汇总。 */
+export function parseTrainingSummary(text: string): TrainingSummary {
+  const emptySummary: TrainingSummary = { epoch: '', trainLoss: '', runtime: '' }
+  const startMatch = /\*{5}\s*train metrics\s*\*{5}/i.exec(text)
+  if (!startMatch) return emptySummary
+
+  const tail = text.slice(startMatch.index + startMatch[0].length)
+  const endMatch = /\*{5}\s*train metrics end\s*\*{5}/i.exec(tail)
+  const body = endMatch ? tail.slice(0, endMatch.index) : tail
+  return {
+    epoch: extractSummaryValue(body, 'epoch'),
+    trainLoss: extractSummaryValue(body, 'train_loss'),
+    runtime: extractSummaryValue(body, 'train_runtime'),
+  }
+}
+
+export function parseTrainingLog(text: string): ParsedTrainingLog {
+  return {
+    metrics: parseTrainingMetrics(text),
+    summary: parseTrainingSummary(text),
+  }
+}
+
+/** 构建单条训练指标曲线。 */
+export function buildMetricChartOption(
+  label: string,
+  data: number[],
+  color: string,
+  logScale = false,
+): EChartsOption {
+  return {
+    grid: { top: 24, right: 20, bottom: 56, left: 56 },
+    tooltip: {
+      trigger: 'axis',
+      axisPointer: { type: 'cross' },
+      backgroundColor: 'rgba(15, 23, 42, 0.9)',
+      borderWidth: 0,
+      textStyle: { color: '#fff', fontSize: 12 },
+    },
+    xAxis: {
+      type: 'category',
+      boundaryGap: false,
+      name: 'Step',
+      nameTextStyle: { color: '#94a3b8', fontSize: 11 },
+      axisLine: { lineStyle: { color: '#e2e8f0' } },
+      axisLabel: { color: '#94a3b8', fontSize: 11 },
+      splitLine: { show: false },
+    },
+    yAxis: {
+      type: logScale ? 'log' : 'value',
+      name: label,
+      nameTextStyle: { color: '#94a3b8', fontSize: 11 },
+      axisLine: { show: false },
+      axisTick: { show: false },
+      axisLabel: { color: '#94a3b8', fontSize: 11 },
+      splitLine: { lineStyle: { color: '#f1f5f9' } },
+    },
+    dataZoom: data.length > 30
+      ? [
+          { type: 'inside', start: 0, end: 100 },
+          { type: 'slider', height: 16, bottom: 8, borderColor: 'transparent', fillerColor: 'rgba(79,70,229,0.08)', handleStyle: { color: '#4f46e5' } },
+        ]
+      : [],
+    series: [
+      {
+        name: label,
+        type: 'line',
+        data,
+        smooth: true,
+        symbol: 'none',
+        lineStyle: { width: 2, color },
+        areaStyle: {
+          color: {
+            type: 'linear',
+            x: 0,
+            y: 0,
+            x2: 0,
+            y2: 1,
+            colorStops: [
+              { offset: 0, color: `${color}55` },
+              { offset: 1, color: `${color}05` },
+            ],
+          },
+        },
+      },
+    ],
+  }
+}