2026-07-21 14:25:38 +08:00
20 changed files with 736 additions and 603 deletions
--- a/frontend/scripts/regression-data-process-wizard.mjs
+++ b/frontend/scripts/regression-data-process-wizard.mjs
@@ -41,15 +41,20 @@ assert.match(viewSource, /onBeforeRouteLeave\(async \(\) =>/, '路由离开确
 assert.doesNotMatch(viewSource, /window\.confirm|ElMessageBox/, '创建页仍在使用系统或 Element Plus 确认框')

 assert.match(viewSource, /const WIZARD_STEPS = \[/, '向导步骤尚未改为固定常量')
-for (const title of ['创建任务', '上传文件', '数据预览', '开始生成', '结果编辑与保存']) {
+for (const title of ['创建任务', '大模型选择', '上传文件', '数据预览', '开始生成', '结果编辑与保存']) {
  assert.ok(viewSource.includes(`title: '${title}'`), `缺少固定步骤：${title}`)
 }
 assert.match(
  viewSource,
-  /\{ id: 'create',[\s\S]*?\{ id: 'upload',[\s\S]*?\{ id: 'preview',[\s\S]*?\{ id: 'generate',[\s\S]*?\{ id: 'results'/,
-  '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',
+  /\{ id: 'create',[\s\S]*?\{ id: 'model',[\s\S]*?\{ id: 'upload',[\s\S]*?\{ id: 'preview',[\s\S]*?\{ id: 'generate',[\s\S]*?\{ id: 'results'/,
+  '六步向导顺序必须为创建任务、大模型选择、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(
+  viewStyleSource,
+  /@media \(max-width: 1100px\)[\s\S]*?\.step-title\s*\{[\s\S]*?display:\s*none[\s\S]*?\.step-item\.is-active \.step-title\s*\{[\s\S]*?display:\s*block/,
+  '六步向导在中等宽度下没有收起非当前步骤标题',
+)
 assert.match(draftSource, /localStorage\.setItem\(DATA_PROCESS_DRAFT_STORAGE_KEY/, '草稿没有持久化')
 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')
@@ -57,6 +62,7 @@ assert.ok(viewSource.split('\n').length < 800, 'DataProcessCreateView 拆分后

 const expectedComponents = [
  'TaskSetupStep.vue',
+  'ModelSelectionStep.vue',
  'SourceUploadStep.vue',
  'PreviewCompareStep.vue',
  'GenerationStep.vue',
@@ -82,7 +88,7 @@ for (const field of ['sourceStart', 'sourceEnd', 'originalContent', 'editedConte
  assert.ok(typesSource.includes(field), `PreviewItem 缺少字段：${field}`)
 }
 assert.match(typesSource, /sourceFileId/, 'PreviewItem 缺少来源文件标识')
-assert.match(typesSource, /export type StepId = 'create' \| 'upload' \| 'preview' \| 'generate' \| 'results'/, '步骤类型缺少独立上传步骤')
+assert.match(typesSource, /export type StepId = 'create' \| 'model' \| 'upload' \| 'preview' \| 'generate' \| 'results'/, '步骤类型缺少独立大模型选择步骤')
 assert.match(modelSource, /export function buildPreviewItems/, '缺少切片来源映射生成函数')
 assert.match(modelSource, /export function sourceLines/, '缺少源文件行偏移生成函数')
 assert.match(modelSource, /sourceFileId/, '切片生成没有写入来源文件标识')
@@ -102,10 +108,10 @@ for (const marker of [
  'preview-editor',
  'scrollIntoView',
 ]) {
-  assert.ok(previewSource.includes(marker), `第三步缺少结构或行为：${marker}`)
+  assert.ok(previewSource.includes(marker), `第四步缺少结构或行为：${marker}`)
 }
-assert.match(previewSource, /sourceStart/, '第三步未使用来源起始偏移')
-assert.match(previewSource, /sourceEnd/, '第三步未使用来源结束偏移')
+assert.match(previewSource, /sourceStart/, '第四步未使用来源起始偏移')
+assert.match(previewSource, /sourceEnd/, '第四步未使用来源结束偏移')
 assert.match(previewSource, /filterable/, '文件选择器必须可搜索')
 assert.match(previewSource, /当前文件/, '预览缺少当前文件切换器')
 assert.doesNotMatch(previewSource, /located-badge|sync-label|已定位到/, '源文件栏不应显示冗余定位提示')
@@ -130,13 +136,14 @@ assert.match(previewSource, />保存修改<\/el-button>/, '编辑器缺少保存
 assert.match(previewSource, /\.editor-actions\s*\{[\s\S]*?justify-content:\s*flex-end/, '取消和保存按钮必须在编辑器右侧对齐')
 assert.doesNotMatch(previewSource, /item-token|item-status|modifiedOnly|仅看已修改/, '切片列表不应再显示 Token 或修改状态')
 assert.match(previewSource, /\.preview-editor\s*\{[\s\S]*?flex:\s*1 1 auto[\s\S]*?overflow-y:\s*auto/, '编辑模式必须占据右侧剩余区域并可滚动')
-assert.match(previewSource, /@media \(max-width: 900px\)/, '第三步缺少窄屏上下布局')
+assert.match(previewSource, /@media \(max-width: 900px\)/, '第四步缺少窄屏上下布局')

 const taskSetupPath = path.join(createDir, 'TaskSetupStep.vue')
 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 modelSelectionPath = path.join(createDir, 'ModelSelectionStep.vue')
 const sourceUploadPath = path.join(createDir, 'SourceUploadStep.vue')
 const [
  taskSetupSource,
@@ -144,6 +151,7 @@ const [
  unstructuredOptionsSource,
  datasetSplitEditorSource,
  generationControlSource,
+  modelSelectionSource,
  sourceUploadSource,
 ] = await Promise.all([
  readFile(taskSetupPath, 'utf8'),
@@ -151,6 +159,7 @@ const [
  readFile(unstructuredOptionsPath, 'utf8'),
  readFile(datasetSplitEditorPath, 'utf8'),
  readFile(generationOptionsPath, 'utf8'),
+  readFile(modelSelectionPath, 'utf8'),
  readFile(sourceUploadPath, 'utf8'),
 ])
 const taskSetupFeatureSource = [
@@ -172,20 +181,29 @@ assert.match(unstructuredOptionsSource, /<DatasetSplitEditor/, '非结构化选
 for (const marker of ['<el-upload', '源数据上传', '数据源配置', 'uploadedFiles']) {
  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, /<ModelSelectionStep\s+[\s\S]*?v-else-if="currentStepId === 'model'"/, '第二步没有挂载独立大模型选择组件')
+assert.match(viewSource, /<SourceUploadStep\s+[\s\S]*?v-else-if="currentStepId === 'upload'"/, '第三步没有挂载独立上传组件')
+assert.match(viewSource, /if \(currentStepId\.value === 'create'\) return '继续：选择大模型'/, '第一步主按钮没有指向大模型选择')
+assert.match(viewSource, /if \(currentStepId\.value === 'model'\) return '继续：上传文件'/, '第二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'/, '第三步主按钮没有指向数据预览')
 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)
+const nextFromModelStart = viewSource.indexOf('async function nextFromModel()', nextFromCreateStart)
+const nextFromUploadStart = viewSource.indexOf('function nextFromUpload()', nextFromModelStart)
 const selectPreviewFileStart = viewSource.indexOf('function selectPreviewFile(', nextFromUploadStart)
-assert.ok(nextFromCreateStart >= 0 && nextFromUploadStart > nextFromCreateStart, '缺少创建步骤与上传步骤的独立跳转函数')
-const nextFromCreateSource = viewSource.slice(nextFromCreateStart, nextFromUploadStart)
+assert.ok(
+  nextFromCreateStart >= 0 && nextFromModelStart > nextFromCreateStart && nextFromUploadStart > nextFromModelStart,
+  '缺少创建、大模型选择与上传步骤的独立跳转函数',
+)
+const nextFromCreateSource = viewSource.slice(nextFromCreateStart, nextFromModelStart)
+const nextFromModelSource = viewSource.slice(nextFromModelStart, nextFromUploadStart)
 const nextFromUploadSource = viewSource.slice(nextFromUploadStart, selectPreviewFileStart)
 assert.match(nextFromCreateSource, /taskSetupRef\.value\?\.validate\(\)/, '创建步骤继续前没有校验任务配置')
-assert.match(nextFromCreateSource, /goToStep\('upload'\)/, '创建步骤校验通过后没有进入上传文件')
+assert.match(nextFromCreateSource, /goToStep\('model'\)/, '创建步骤校验通过后没有进入大模型选择')
 assert.doesNotMatch(nextFromCreateSource, /uploadedFiles|buildPreviewItems/, '创建步骤仍在校验文件或提前生成预览')
+assert.match(nextFromModelSource, /modelSelectionRef\.value\?\.validate\(\)/, '大模型选择步骤继续前没有校验模型配置')
+assert.match(nextFromModelSource, /goToStep\('upload'\)/, '大模型选择完成后没有进入上传文件')
 assert.match(nextFromUploadSource, /uploadedFiles\.value\.length === 0/, '上传步骤继续前没有校验源数据')
 assert.match(nextFromUploadSource, /buildPreviewItems\(/, '上传步骤没有在进入预览前生成预览数据')
 assert.match(nextFromUploadSource, /goToStep\('preview'\)/, '上传步骤完成后没有进入数据预览')
@@ -217,8 +235,6 @@ for (const option of [
  assert.ok(structuredOptionsSource.includes(option), `结构化预处理缺少选项：${option}`)
 }
 assert.ok(structuredOptionsSource.includes('生成选项'), '结构化配置缺少生成选项分类')
-assert.ok(structuredOptionsSource.includes('语义丰富表达'), '生成选项缺少语义丰富表达开关')
-assert.ok(structuredOptionsSource.includes('使用大模型将问答表述得更自然、柔和'), '语义丰富表达缺少辅助说明')
 for (const splitName of ['训练集', '验证集', '测试集']) {
  assert.ok(datasetSplitEditorSource.includes(splitName), `生成选项缺少数据集划分：${splitName}`)
 }
@@ -232,7 +248,6 @@ for (const splitField of ['train', 'validation', 'test']) {
    `数据集划分字段 ${splitField} 缺少 0～100 的整数限制`,
  )
 }
-assert.match(structuredOptionsSource, /<el-switch[\s\S]*options\.semanticEnrichment/, '语义丰富表达必须使用开关控件')
 assert.match(structuredOptionsSource, /<el-input-number[\s\S]*options\.qaPairsPerRow[\s\S]*:min="1"[\s\S]*:max="5"/, '每行生成数量必须限制在 1 到 5')
 assert.match(viewSource, /const structuredOptions = ref<StructuredProcessOptions>/, '父页面缺少结构化配置状态')
 assert.match(stateSource, /datasetSplit:\s*\{ train: 80, validation: 10, test: 10 \}/, '数据集划分默认值必须为 80/10/10')
@@ -252,12 +267,16 @@ for (const field of [
  assert.ok(typesSource.includes(field), `生成控制配置缺少字段：${field}`)
  assert.ok(implementationSource.includes(field), `父页面默认值或草稿状态缺少字段：${field}`)
 }
-assert.match(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(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(structuredOptionsSource, /GenerationOptionsPanel/, '结构化生成选项没有复用统一的质量筛选组件')
+assert.match(unstructuredOptionsSource, /GenerationOptionsPanel/, '非结构化生成选项没有复用统一的质量筛选组件')
 assert.match(structuredOptionsSource, /:options="options"/, '结构化生成选项未接入统一配置组件')
 assert.match(unstructuredOptionsSource, /:options="options"/, '非结构化生成选项未接入统一配置组件')
-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(taskSetupFeatureSource, /section="model"/, '独立大模型分类没有挂载模型配置')
+assert.doesNotMatch(taskSetupFeatureSource, /<h3>大模型<\/h3>|section="model"/, '第一步不应继续承载大模型配置')
+assert.match(modelSelectionSource, /<h3[^>]*>大模型选择<\/h3>/, '独立步骤缺少大模型选择标题')
+assert.match(modelSelectionSource, /section="model"/, '独立步骤没有挂载模型配置')
+assert.match(modelSelectionSource, /defineExpose\(\{ validate \}\)/, '独立大模型选择步骤没有暴露继续前校验')
+assert.match(modelSelectionSource, /class="form-section"/, '大模型选择步骤没有沿用第一步的通栏表单分区')
+assert.doesNotMatch(modelSelectionSource, /max-width:\s*980px/, '大模型选择步骤不应使用比第一步更窄的固定内容宽度')
 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']) {
@@ -267,15 +286,21 @@ assert.match(generationControlSource, /filterable/, '数据生成模型下拉必
 assert.match(generationControlSource, /maxlength="500"/, '默认提示语缺少合理的长度限制')
 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(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.match(generationControlSource, /\.model-config-group\s*\{[\s\S]*?padding:\s*0[\s\S]*?border:\s*0/, '独立大模型步骤仍存在嵌套卡片挤压')
+assert.match(
+  generationControlSource,
+  /\.model-config-group \.advanced-settings-grid\s*\{[\s\S]*?grid-template-columns:\s*1fr/,
+  '大模型高级参数没有改为与第一步一致的纵向布局',
+)
 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')
-assert.match(taskSetupSource, /generationValidationMessage/, '生成模型与质量规则缺少继续前校验')
+assert.match(taskSetupSource, /qualityValidationMessage/, '质量规则缺少继续前校验')
 assert.match(viewSource, /useModelsStore/, '创建页没有加载模型列表')
 assert.match(viewSource, /model\.type === 'LLM'/, '数据生成模型列表没有排除非大模型')
-assert.match(viewSource, /:generation-models="generationModels"/, '创建页没有向生成选项传递模型列表')
+assert.match(viewSource, /<ModelSelectionStep[\s\S]*?:models="generationModels"/, '创建页没有向独立大模型选择步骤传递模型列表')

 assert.match(typesSource, /export interface UnstructuredProcessOptions/, '缺少非结构化处理选项类型')
 for (const field of [
@@ -312,11 +337,7 @@ for (const method of ['自动语义切分', '按标题和段落', '按固定长
 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(unstructuredOptionsSource.includes(label), `切分选项缺少配置：${label}`)
 }
-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(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(unstructuredOptionsSource, /v-if="advancedChunkSettingsOpen"/, '高级设置内容没有默认收起')
-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(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.doesNotMatch(unstructuredOptionsSource, /advancedChunkSettingsOpen|>高级设置</, '切分核心参数不应再隐藏在高级设置中')
 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(unstructuredOptionsSource, /const preserveSpecialContentEnabled = computed/, '特殊内容保护没有合并为单一开关')
 assert.match(unstructuredOptionsSource, /function updateSpecialContentProtection/, '特殊内容保护开关缺少更新逻辑')
@@ -324,7 +345,7 @@ assert.match(unstructuredOptionsSource, /options\.chunkSize[\s\S]*?:min="200"[\s
 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']) {
+for (const label of ['每个切片生成数量', '数据集划分']) {
  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']) {
@@ -862,4 +883,4 @@ assert.match(
 )
 assert.match(previewSource, /height:\s*clamp\(560px,\s*calc\(100vh - 370px\),\s*720px\)/, '对照预览高度不足以展示切片正文')

-console.log('数据处理五步向导回归检查通过')
+console.log('数据处理六步向导回归检查通过')
--- a/frontend/scripts/regression-eval-create-wizard.mjs
+++ b/frontend/scripts/regression-eval-create-wizard.mjs
@@ -11,6 +11,8 @@ const [
  createSource,
  taskStepSource,
  ruleStepSource,
+  basicMetricStepSource,
+  startStepSource,
  dimensionFieldsSource,
  listSource,
  routerSource,
@@ -19,6 +21,8 @@ const [
  readFile(path.join(evalViewDir, 'EvalCreateView.vue'), 'utf8'),
  readFile(path.join(evalViewDir, 'create/EvalTaskSetupStep.vue'), 'utf8'),
  readFile(path.join(evalViewDir, 'create/EvalRuleSetupStep.vue'), 'utf8'),
+  readFile(path.join(evalViewDir, 'create/BasicMetricSetupStep.vue'), 'utf8'),
+  readFile(path.join(evalViewDir, 'create/StartEvalStep.vue'), 'utf8'),
  readFile(path.join(evalViewDir, 'create/DimensionFormFields.vue'), 'utf8'),
  readFile(path.join(evalViewDir, 'EvalView.vue'), 'utf8'),
  readFile(path.join(sourceRoot, 'router/index.ts'), 'utf8'),
@@ -74,21 +78,43 @@ assert.match(

 assert.match(createSource, /EvalTaskSetupStep/, '向导缺少任务配置子组件')
 assert.match(createSource, /EvalRuleSetupStep/, '向导缺少评测规则子组件')
+assert.match(createSource, /BasicMetricSetupStep/, '向导缺少基础评测指标子组件')
+assert.match(createSource, /StartEvalStep/, '向导缺少开始评测确认子组件')
 assert.match(
  createSource,
-  /任务配置[\s\S]*?评测规则/,
-  '评测创建向导必须按“任务配置、评测规则”定义两个步骤',
+  /任务配置[\s\S]*?大模型评测指标[\s\S]*?基础评测指标[\s\S]*?开始评测/,
+  '评测创建向导必须按“任务配置、大模型评测指标、基础评测指标、开始评测”定义四个步骤',
 )
 assert.match(
  ruleStepSource,
  /:show-description=["']false["']/,
-  '新建评测任务第二步不应显示重复的规则描述输入框',
+  '大模型评测指标步骤不应显示重复的规则描述输入框',
+)
+assert.match(
+  ruleStepSource,
+  /:show-status-settings=["']false["']/,
+  '大模型评测指标步骤不应显示维度管理状态字段',
+)
+assert.match(
+  ruleStepSource,
+  /LLM_METRIC_TYPES[^=]*=\s*\[['"]classification['"],\s*['"]metric['"]\][\s\S]*?:allowed-types=["']LLM_METRIC_TYPES["']/,
+  '大模型评测指标步骤只应提供依赖大模型的分类与评分指标',
 )
 assert.match(
  dimensionFieldsSource,
  /v-if=["']showDescription["'][^>]*label=["']描述["']/,
  '共享维度字段必须支持按场景隐藏描述输入框',
 )
+assert.match(
+  dimensionFieldsSource,
+  /score_min[\s\S]*?score_max[\s\S]*?pass_threshold[\s\S]*?Math\.min[\s\S]*?Math\.max/,
+  '指标型评分区间变化时必须把通过阈值约束在合法范围内',
+)
+assert.match(
+  dimensionFieldsSource,
+  /prop=["']score_min["'][\s\S]*?:max=["'][^"']*score_max[^"']*["'][\s\S]*?prop=["']score_max["'][\s\S]*?:min=["'][^"']*score_min[^"']*["']/,
+  '评分最小值和最大值输入必须具备交叉边界约束',
+)
 assert.match(createSource, /currentStep\s*=\s*ref\(0\)/, '评测创建向导缺少当前步骤状态')
 assert.match(createSource, /class=["']custom-wizard-steps["']/, '评测向导未复用数据处理的自定义步骤条结构')
 assert.match(createSource, /class=["']step-connector["']/, '评测向导步骤条缺少连接线')
@@ -102,8 +128,18 @@ assert.match(
 )
 assert.match(
  createSource,
-  /<EvalRuleSetupStep[\s\S]*?v-else/,
-  '评测规则子组件没有绑定第二步',
+  /<EvalRuleSetupStep[\s\S]*?currentStep[^"']*1/,
+  '大模型评测指标子组件没有绑定第二步',
+)
+assert.match(
+  createSource,
+  /<BasicMetricSetupStep[\s\S]*?currentStep[^"']*2/,
+  '基础评测指标子组件没有绑定第三步',
+)
+assert.match(
+  createSource,
+  /<StartEvalStep[\s\S]*?v-else/,
+  '开始评测确认子组件没有绑定第四步',
 )

 assert.match(taskStepSource, /prop=["']dataset_id["']/, '数据集字段缺少表单校验标识')
@@ -113,16 +149,15 @@ assert.match(
  '选择评测数据集时必须校验 dataset_id',
 )

-assert.match(createSource, /['"]baseline['"]/, '向导缺少基线评测分支')
 assert.match(
-  `${createSource}\n${ruleStepSource}`,
-  /已有维度|existing/,
-  '评测规则步骤缺少使用已有维度的分支',
+  basicMetricStepSource,
+  /BLEU[\s\S]*?ROUGE[\s\S]*?(?:Cosine|余弦)/,
+  '基础评测指标步骤必须提供 BLEU、ROUGE 与余弦相似度配置',
 )
 assert.match(
-  `${createSource}\n${ruleStepSource}`,
-  /新建(?:评测)?(?:维度|规则)|create/,
-  '评测规则步骤缺少新建维度的分支',
+  basicMetricStepSource,
+  /bleu_enabled[\s\S]*?rouge_enabled[\s\S]*?cosine_enabled/,
+  '基础评测指标必须支持分别决定是否启用',
 )

 assert.match(
@@ -137,17 +172,18 @@ assert.match(
 )
 assert.match(
  createSource,
-  /eval_type[\s\S]*?baseline[\s\S]*?(?:dimension_id|createDimension)/,
-  '基线评测必须有明确的不创建自定义维度分支',
+  /basic_metrics[\s\S]*?bleu[\s\S]*?rouge[\s\S]*?cosine/,
+  '启动评测时必须提交基础评测指标配置',
 )
 assert.match(
-  createSource,
-  /ruleMode[\s\S]*?['"]existing['"]/,
-  '已有维度模式必须有独立的提交分支',
+  startStepSource,
+  /任务信息[\s\S]*?大模型评测指标[\s\S]*?基础评测指标/,
+  '开始评测步骤必须展示前三步配置摘要',
 )

 assert.match(createSource, /:loading=["']submitting["']/, '最终提交按钮缺少 loading 状态')
 assert.match(createSource, /:disabled=["'][^"']*submitting/, '提交期间必须禁用重复操作')
+assert.match(createSource, />\s*开始评测\s*</, '最终提交按钮必须命名为“开始评测”')
 assert.match(createSource, /class=["'][^"']*wizard-footer/, '评测向导缺少统一底部操作栏')
 assert.match(
  createSource,
--- a/frontend/src/api/modules/eval.ts
+++ b/frontend/src/api/modules/eval.ts
@@ -1,5 +1,5 @@
 import { get, post, put, del } from '../request'
-import type { EvalTask, EvalTaskDetail, Dimension } from '@/types'
+import type { EvalTask, EvalTaskDetail, Dimension, StartEvalPayload } from '@/types'

 /** 评测任务列表 */
 export const getEvalList = () => get<EvalTask[]>('/model-eval')
@@ -11,7 +11,7 @@ export const getEvalDetail = (id: string | number) => get<EvalTaskDetail>(`/mode
 export const deleteEval = (id: string | number) => del(`/model-eval/${id}`)

 /** 启动评测 */
-export const startEval = (data: any) => post('/model-eval/start', data)
+export const startEval = (data: StartEvalPayload) => post('/model-eval/start', data)

 /** 评测维度列表 */
 export const getDimensionList = () => get<Dimension[]>('/dimension')
--- a/frontend/src/components/AppSidebar.vue
+++ b/frontend/src/components/AppSidebar.vue
@@ -36,7 +36,7 @@ const menuGroups = [
  {
    title: '模型服务',
    items: [
-      { key: 'fine-tune', label: '模型微调', icon: 'fa-cogs', to: '/fine-tune' },
+      { key: 'fine-tune', label: '模型训练', icon: 'fa-cogs', to: '/fine-tune' },
      { key: 'model-eval', label: '模型评测', icon: 'fa-line-chart', to: '/model-eval' },
      { key: 'model-inference', label: '模型推理', icon: 'fa-server', to: '/model-inference' },
      { key: 'model-manage', label: '模型管理', icon: 'fa-cube', to: '/model-manage' },
@@ -45,8 +45,8 @@ const menuGroups = [
  {
    title: '数据治理',
    items: [
-      { key: 'data-process', label: '数据处理', icon: 'fa-filter', to: '/data-process' },
      { key: 'dataset', label: '数据集管理', icon: 'fa-file-text', to: '/dataset' },
+      { key: 'data-process', label: '数据处理', icon: 'fa-filter', to: '/data-process' },
    ],
  },
  {
--- a/frontend/src/router/index.ts
+++ b/frontend/src/router/index.ts
@@ -25,7 +25,7 @@ const routes: RouteRecordRaw[] = [
        path: 'fine-tune',
        name: 'fine-tune',
        component: () => import('@/views/fine-tune/FineTuneListView.vue'),
-        meta: { title: '模型微调', pageSurface: 'self' },
+        meta: { title: '模型训练', pageSurface: 'self' },
      },
      {
        path: 'fine-tune/create',
--- a/frontend/src/types/index.ts
+++ b/frontend/src/types/index.ts
@@ -224,6 +224,34 @@ export interface EvalTask {
  create_time?: string
 }

+/** 启动评测时提交的可选基础指标配置。 */
+export interface BasicEvalMetricsConfig {
+  bleu: {
+    enabled: boolean
+    ngram: number
+  }
+  rouge: {
+    enabled: boolean
+    methods: string[]
+  }
+  cosine: {
+    enabled: boolean
+  }
+  output_precision: number
+}
+
+export interface StartEvalPayload {
+  eval_task_name: string
+  eval_type: EvalType
+  model_id: string | number
+  gpu_id: string | number
+  dataset_id: string | number
+  dimension_id: string | number
+  data_source: 'dataset' | 'inference'
+  leaderboard: boolean
+  basic_metrics: BasicEvalMetricsConfig
+}
+
 /** 单个测试样本的评测结果 */
 export interface EvalSampleResult {
  id: number | string
--- a/frontend/src/views/data-process/DataProcessCreateView.vue
+++ b/frontend/src/views/data-process/DataProcessCreateView.vue
@@ -5,6 +5,7 @@ import { ElMessage, type UploadFile } from 'element-plus'
 import { storeToRefs } from 'pinia'
 import AppConfirmDialog from '@/components/AppConfirmDialog.vue'
 import TaskSetupStep from './create/TaskSetupStep.vue'
+import ModelSelectionStep from './create/ModelSelectionStep.vue'
 import SourceUploadStep from './create/SourceUploadStep.vue'
 import PreviewCompareStep from './create/PreviewCompareStep.vue'
 import GenerationStep from './create/GenerationStep.vue'
@@ -22,6 +23,7 @@ import { useDataProcessGeneration } from './create/useDataProcessGeneration'
 import { useModelsStore } from '@/stores/models'
 import type {
  ExternalDataSource,
+  GenerationControlOptions,
  PreviewItem,
  ProcessType,
  StepId,
@@ -35,22 +37,27 @@ const modelsStore = useModelsStore()
 const { list: modelList } = storeToRefs(modelsStore)
 const generationModels = computed(() => modelList.value.filter((model) => model.type === 'LLM'))
 const taskSetupRef = ref<InstanceType<typeof TaskSetupStep>>()
+const modelSelectionRef = ref<InstanceType<typeof ModelSelectionStep>>()
 const confirmDialogRef = ref<InstanceType<typeof AppConfirmDialog>>()
 const PREVIEW_MODEL_VERSION = 'document-chunk-v2'

 const WIZARD_STEPS = [
  { id: 'create', title: '创建任务', desc: '填写任务信息与处理配置' },
+  { id: 'model', title: '大模型选择', desc: '选择生成模型并设置输出要求' },
  { id: 'upload', title: '上传文件', desc: '上传或接入待处理的源数据' },
  { id: 'preview', title: '数据预览', desc: '核对源文件与预览内容' },
  { 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' },
+  { id: 'results', title: '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 currentStep = ref(0)
 const currentStepId = computed<StepId>(() => WIZARD_STEPS[currentStep.value]?.id ?? 'create')
 const task = reactive({ name: '', description: '' })
-const processType = ref<ProcessType>('unstructured')
+const processType = ref<ProcessType>('structured')
 const structuredOptions = ref<StructuredProcessOptions>(createDefaultStructuredOptions())
 const unstructuredOptions = ref<UnstructuredProcessOptions>(createDefaultUnstructuredOptions())
+const modelSelectionOptions = computed<GenerationControlOptions>(() => (
+  processType.value === 'unstructured' ? unstructuredOptions.value : structuredOptions.value
+))
 const uploadedFiles = ref<UploadedDataFile[]>([])

 const externalSource = reactive<ExternalDataSource>({
@@ -119,7 +126,8 @@ const previewFiles = computed(() => uploadedFiles.value.map((file) => {
  }
 }))
 const primaryActionLabel = computed(() => {
-  if (currentStepId.value === 'create') return '继续：上传文件'
+  if (currentStepId.value === 'create') return '继续：选择大模型'
+  if (currentStepId.value === 'model') return '继续：上传文件'
  if (currentStepId.value === 'upload') return '继续：数据预览'
  if (currentStepId.value === 'preview') return '确认预览并继续'
  if (currentStepId.value === 'results') return '保存任务'
@@ -144,6 +152,14 @@ function goToStep(stepId: StepId) {
  if (nextStepIndex >= 0) currentStep.value = nextStepIndex
 }

+function updateModelSelectionOptions(value: GenerationControlOptions) {
+  if (processType.value === 'unstructured') {
+    unstructuredOptions.value = { ...unstructuredOptions.value, ...value }
+    return
+  }
+  structuredOptions.value = { ...structuredOptions.value, ...value }
+}
+
 const { persistDraft, restoreDraft } = useDataProcessDraft({
  currentStepId,
  task,
@@ -393,6 +409,12 @@ function resetSourceDataForProcessTypeChange() {
 async function nextFromCreate() {
  const valid = await taskSetupRef.value?.validate()
  if (!valid) return
+  goToStep('model')
+}
+
+async function nextFromModel() {
+  const valid = await modelSelectionRef.value?.validate()
+  if (!valid) return
  goToStep('upload')
 }

@@ -504,6 +526,10 @@ async function handlePrimaryAction() {
    await nextFromCreate()
    return
  }
+  if (currentStepId.value === 'model') {
+    await nextFromModel()
+    return
+  }
  if (currentStepId.value === 'upload') {
    nextFromUpload()
    return
@@ -596,9 +622,13 @@ onMounted(() => {
      <div class="wizard-main-inner">
        <div class="wizard-steps-container">
          <div class="custom-wizard-steps">
+            <template v-for="(step, index) in WIZARD_STEPS" :key="step.id">
+              <div 
+                v-if="index !== 0" 
+                class="step-connector"
+                :class="{ 'is-active': currentStep >= index }"
+              ></div>
              <div 
-              v-for="(step, index) in WIZARD_STEPS" 
-              :key="step.id"
                class="step-item"
                :class="{ 
                  'is-active': currentStep === index, 
@@ -606,7 +636,6 @@ onMounted(() => {
                }"
                :aria-current="currentStep === index ? 'step' : undefined"
              >
-              <div v-if="index !== 0" class="step-connector"></div>
                <div class="step-node">
                  <div class="step-icon">
                    <i v-if="currentStep > index" class="fa fa-check" />
@@ -617,6 +646,7 @@ onMounted(() => {
                  </div>
                </div>
              </div>
+            </template>
          </div>
        </div>

@@ -629,7 +659,14 @@ onMounted(() => {
            v-model:process-type="processType"
            v-model:structured-options="structuredOptions"
            v-model:unstructured-options="unstructuredOptions"
-            :generation-models="generationModels"
+          />
+
+          <ModelSelectionStep
+            v-else-if="currentStepId === 'model'"
+            ref="modelSelectionRef"
+            :options="modelSelectionOptions"
+            :models="generationModels"
+            @update:options="updateModelSelectionOptions"
          />

          <SourceUploadStep
--- a/frontend/src/views/data-process/DataProcessListView.vue
+++ b/frontend/src/views/data-process/DataProcessListView.vue
@@ -92,7 +92,7 @@ function formatDateTime(value?: string) {
      :data="dataList"
      searchable
      :search-fields="['name']"
-      create-text="新建数据处理任务"
+      create-text="新建数据处理"
      create-to="/data-process/create"
      row-key="id"
      :page-size="10"
--- a/frontend/src/views/data-process/create/DatasetSplitEditor.vue
+++ b/frontend/src/views/data-process/create/DatasetSplitEditor.vue
@@ -128,19 +128,19 @@ function ariaLabel(label: string) {
 }

 .dataset-split-row {
-  align-items: stretch;
-  flex-direction: column;
+  align-items: flex-start;
 }

 .dataset-split-config {
  display: grid;
  gap: 10px;
+  flex: 0 0 460px;
 }

 .dataset-split-grid {
  display: grid;
  grid-template-columns: repeat(3, minmax(0, 1fr));
-  gap: 16px;
+  gap: 12px;
 }

 .dataset-split-field {
@@ -173,6 +173,15 @@ function ariaLabel(label: string) {
 }

@media (max-width: 900px) {
+  .dataset-split-row {
+    flex-direction: column;
+    align-items: stretch;
+  }
+  
+  .dataset-split-config {
+    flex: auto;
+  }
+
  .dataset-split-grid {
    grid-template-columns: minmax(0, 1fr);
  }
--- a/frontend/src/views/data-process/create/GenerationOptionsPanel.vue
+++ b/frontend/src/views/data-process/create/GenerationOptionsPanel.vue
@@ -5,13 +5,11 @@ import type { GenerationControlOptions } from './types'

 const props = defineProps<{
  options: GenerationControlOptions
-  models: ModelItem[]
+  models?: ModelItem[]
  section: 'model' | 'quality'
  validationMessage?: string
 }>()

-const advancedModelSettingsOpen = ref(false)
-
 const emit = defineEmits<{
  'update:options': [value: GenerationControlOptions]
 }>()
@@ -61,7 +59,7 @@ function sectionValidationMessage() {
        @update:model-value="updateField('generationModelId', $event)"
      >
        <el-option
-          v-for="model in models"
+          v-for="model in models || []"
          :key="model.id"
          :label="model.name"
          :value="model.id"
@@ -99,20 +97,6 @@ function sectionValidationMessage() {
      </div>
    </div>

-    <button
-      type="button"
-      class="advanced-settings-toggle"
-      :aria-expanded="advancedModelSettingsOpen"
-      @click="advancedModelSettingsOpen = !advancedModelSettingsOpen"
-    >
-      <span>
-        <strong>高级设置</strong>
-        <small>温度、最大输出长度与格式约束</small>
-      </span>
-      <i class="fa" :class="advancedModelSettingsOpen ? 'fa-chevron-up' : 'fa-chevron-down'" />
-    </button>
-
-    <div v-if="advancedModelSettingsOpen" class="advanced-settings-panel">
    <div class="advanced-settings-grid">
      <label class="config-field">
        <span class="config-field-label">生成温度 (Temperature)</span>
@@ -157,7 +141,6 @@ function sectionValidationMessage() {
        @update:model-value="updateField('jsonMode', Boolean($event))"
      />
    </div>
-    </div>

    <p v-if="sectionValidationMessage()" class="generation-validation-message" role="alert">
      {{ sectionValidationMessage() }}
@@ -251,7 +234,22 @@ function sectionValidationMessage() {
 .model-config-group {
  display: flex;
  flex-direction: column;
-  gap: 16px;
+  gap: 20px;
+  padding: 0;
+  border: 0;
+  border-radius: 0;
+  background: transparent;
+}
+
+.model-config-group .advanced-settings-grid {
+  grid-template-columns: 1fr;
+  gap: 20px;
+}
+
+.model-config-group .config-field {
+  padding: 0;
+  border: 0;
+  border-radius: 0;
 }

 .model-field {
@@ -297,61 +295,6 @@ function sectionValidationMessage() {
  }
 }

-.advanced-settings-toggle {
-  display: flex;
-  width: 100%;
-  min-height: 48px;
-  align-items: center;
-  justify-content: space-between;
-  gap: 16px;
-  margin-top: 4px;
-  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;
-  }
-
-  > 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: 16px;
-  margin-top: -8px;
-  padding: 16px;
-  background: #fff;
-  border: 1px solid #e2e5ec;
-  border-radius: 8px;
-}
-
 .advanced-settings-grid {
  display: grid;
  grid-template-columns: repeat(2, minmax(0, 1fr));
@@ -401,8 +344,8 @@ function sectionValidationMessage() {
  display: flex;
  align-items: center;
  justify-content: space-between;
-  gap: 20px;
-  padding: 14px;
+  gap: 16px;
+  padding: 12px 14px;
  background: #fff;
  border: 1px solid #e2e5ec;
  border-radius: 8px;
--- a/frontend/src/views/data-process/create/ModelSelectionStep.vue
+++ b/frontend/src/views/data-process/create/ModelSelectionStep.vue
@@ -0,0 +1,90 @@
+<script setup lang="ts">
+import { computed, ref } from 'vue'
+import { ElMessage } from 'element-plus'
+import type { ModelItem } from '@/types'
+import type { GenerationControlOptions } from './types'
+import GenerationOptionsPanel from './GenerationOptionsPanel.vue'
+
+const props = defineProps<{
+  options: GenerationControlOptions
+  models: ModelItem[]
+}>()
+
+const emit = defineEmits<{
+  'update:options': [value: GenerationControlOptions]
+}>()
+
+const validationAttempted = ref(false)
+const modelValidationMessage = computed(() => (
+  props.options.generationModelId === '' ? '请选择数据生成模型' : ''
+))
+
+function validate() {
+  validationAttempted.value = true
+  if (!modelValidationMessage.value) return true
+  ElMessage.error(modelValidationMessage.value)
+  return false
+}
+
+defineExpose({ validate })
+</script>
+
+<template>
+  <section class="model-selection-step" aria-labelledby="model-selection-title">
+    <div class="form-section">
+      <div class="section-title-row">
+        <div>
+          <h3 id="model-selection-title">大模型选择</h3>
+          <p>选择本次数据生成使用的模型，并设置统一的输出要求</p>
+        </div>
+      </div>
+
+      <div class="model-form-content">
+        <GenerationOptionsPanel
+          :options="options"
+          :models="models"
+          section="model"
+          :validation-message="validationAttempted ? modelValidationMessage : ''"
+          @update:options="emit('update:options', $event)"
+        />
+      </div>
+    </div>
+  </section>
+</template>
+
+<style scoped lang="scss">
+.model-selection-step {
+  width: 100%;
+}
+
+.form-section {
+  padding: 0 0 26px;
+  margin-bottom: 26px;
+  border-bottom: 1px solid #edf0f5;
+}
+
+.section-title-row {
+  display: flex;
+  align-items: flex-start;
+  justify-content: space-between;
+  gap: 16px;
+
+  h3 {
+    margin: 0 0 5px;
+    color: #2f3747;
+    font-size: 15px;
+    font-weight: 650;
+  }
+
+  p {
+    margin: 0;
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.6;
+  }
+}
+
+.model-form-content {
+  margin-top: 16px;
+}
+</style>
--- a/frontend/src/views/data-process/create/StructuredOptionsPanel.vue
+++ b/frontend/src/views/data-process/create/StructuredOptionsPanel.vue
@@ -1,5 +1,4 @@
 <script setup lang="ts">
-import type { ModelItem } from '@/types'
 import type {
  GenerationControlOptions,
  PreprocessOption,
@@ -10,9 +9,8 @@ import GenerationOptionsPanel from './GenerationOptionsPanel.vue'

 const props = defineProps<{
  options: StructuredProcessOptions
-  generationModels: ModelItem[]
  validationAttempted: boolean
-  generationValidationMessage: string
+  qualityValidationMessage: string
 }>()

 const emit = defineEmits<{
@@ -89,24 +87,11 @@ function updatePreprocessOptions(value: Array<string | number | boolean>) {
    <div class="generation-option-list">
      <GenerationOptionsPanel
        :options="options"
-        :models="generationModels"
        section="quality"
-        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        :validation-message="validationAttempted ? qualityValidationMessage : ''"
        @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>
@@ -128,23 +113,6 @@ function updatePreprocessOptions(value: Array<string | number | boolean>) {
    </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">
@@ -191,6 +159,8 @@ function updatePreprocessOptions(value: Array<string | number | boolean>) {
 }

 .preprocess-option {
+  display: flex;
+  gap: 10px;
  width: 100%;
  min-height: 64px;
  margin-right: 0;
--- a/frontend/src/views/data-process/create/TaskSetupStep.vue
+++ b/frontend/src/views/data-process/create/TaskSetupStep.vue
@@ -1,7 +1,6 @@
 <script setup lang="ts">
 import { computed, ref } from 'vue'
-import type { FormInstance, FormRules } from 'element-plus'
-import type { ModelItem } from '@/types'
+import { ElMessage, type FormInstance, type FormRules } from 'element-plus'
 import type {
  GenerationControlOptions,
  ProcessType,
@@ -17,7 +16,6 @@ const props = defineProps<{
  processType: ProcessType
  structuredOptions: StructuredProcessOptions
  unstructuredOptions: UnstructuredProcessOptions
-  generationModels: ModelItem[]
 }>()

 const emit = defineEmits<{
@@ -52,10 +50,9 @@ const activeGenerationOptions = computed<GenerationControlOptions | null>(() =>
  return null
 })

-const generationValidationMessage = computed(() => {
+const qualityValidationMessage = computed(() => {
  const options = activeGenerationOptions.value
  if (!options) return ''
-  if (options.generationModelId === '') return '请选择数据生成模型'
  if (options.qualityFilterEnabled && !options.filterLowQuality && !options.filterShortContent) {
    return '开启质量筛选后，请至少选择一项筛选规则'
  }
@@ -97,17 +94,28 @@ async function validate() {
  if (!formRef.value) return false
  try {
    await formRef.value.validate()
-    if (props.processType === 'structured' && splitTotal.value !== 100) return false
+    if (props.processType === 'structured' && splitTotal.value !== 100) {
+      ElMessage.error('结构化数据处理：数据集划分比例总和必须为 100%')
+      return false
+    }
    if (props.processType === 'unstructured') {
-      if (unstructuredSplitTotal.value !== 100) return false
+      if (unstructuredSplitTotal.value !== 100) {
+        ElMessage.error('非结构化数据处理：数据集划分比例总和必须为 100%')
+        return false
+      }
      if (chunkValidationMessage.value) {
        unstructuredOptionsPanelRef.value?.revealValidation()
+        ElMessage.error(chunkValidationMessage.value)
        return false
      }
    }
-    if (generationValidationMessage.value) return false
+    if (qualityValidationMessage.value) {
+      ElMessage.error(qualityValidationMessage.value)
+      return false
+    }
    return true
  } catch {
+    ElMessage.error('请完善标红的必填项')
    return false
  }
 }
@@ -117,7 +125,7 @@ defineExpose({ validate })

 <template>
  <section class="task-setup-step">
-    <el-form ref="formRef" :model="formModel" :rules="rules" label-position="top">
+    <el-form ref="formRef" :model="formModel" :rules="rules" label-position="top" scroll-to-error>
      <div class="form-section">
        <h3>基本信息</h3>
        <div class="basic-grid">
@@ -199,9 +207,8 @@ defineExpose({ validate })
      <StructuredOptionsPanel
        v-if="processType === 'structured'"
        :options="structuredOptions"
-        :generation-models="generationModels"
        :validation-attempted="validationAttempted"
-        :generation-validation-message="generationValidationMessage"
+        :quality-validation-message="qualityValidationMessage"
        @update:options="emit('update:structuredOptions', $event)"
      />

@@ -209,9 +216,8 @@ defineExpose({ validate })
        v-if="processType === 'unstructured'"
        ref="unstructuredOptionsPanelRef"
        :options="unstructuredOptions"
-        :generation-models="generationModels"
        :validation-attempted="validationAttempted"
-        :generation-validation-message="generationValidationMessage"
+        :quality-validation-message="qualityValidationMessage"
        :chunk-validation-message="chunkValidationMessage"
        @update:options="emit('update:unstructuredOptions', $event)"
      />
--- a/frontend/src/views/data-process/create/UnstructuredOptionsPanel.vue
+++ b/frontend/src/views/data-process/create/UnstructuredOptionsPanel.vue
@@ -1,6 +1,5 @@
 <script setup lang="ts">
 import { computed, ref, watch } from 'vue'
-import type { ModelItem } from '@/types'
 import type {
  ChunkMethod,
  GenerationControlOptions,
@@ -12,9 +11,8 @@ import GenerationOptionsPanel from './GenerationOptionsPanel.vue'

 const props = defineProps<{
  options: UnstructuredProcessOptions
-  generationModels: ModelItem[]
  validationAttempted: boolean
-  generationValidationMessage: string
+  qualityValidationMessage: string
  chunkValidationMessage: string
 }>()

@@ -47,8 +45,6 @@ const UNSTRUCTURED_NUMBER_LIMITS = {

 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),
 ))
@@ -61,14 +57,6 @@ const preserveSpecialContentEnabled = computed(() => (
  && 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],
@@ -117,7 +105,7 @@ function updateUnstructuredNumber(field: UnstructuredNumberField, value: number
 }

 function revealValidation() {
-  advancedChunkSettingsOpen.value = true
+  // Advanced settings are now flattened, no need to open toggle
 }

 defineExpose({ revealValidation })
@@ -131,35 +119,39 @@ defineExpose({ revealValidation })
        <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
+    <div class="preprocess-option-grid">
+      <label class="preprocess-option" :class="{ 'is-checked': smartPreprocessEnabled }">
+        <el-checkbox
          :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
+        <span class="preprocess-option-copy">
+          <strong>智能预处理</strong>
+          <small>自动清理、解析、去重及保留上下文</small>
+        </span>
+      </label>
+
+      <label class="preprocess-option" :class="{ 'is-checked': desensitizeEnabled }">
+        <el-checkbox
          :model-value="desensitizeEnabled"
-          aria-label="敏感信息脱敏"
-          inline-prompt
-          active-text="开"
-          inactive-text="关"
          @update:model-value="updateDesensitize"
        />
-      </div>
+        <span class="preprocess-option-copy">
+          <strong>敏感信息脱敏</strong>
+          <small>处理姓名、手机号等隐私信息</small>
+        </span>
+      </label>
+
+      <label class="preprocess-option" :class="{ 'is-checked': preserveSpecialContentEnabled }">
+        <el-checkbox
+          :model-value="preserveSpecialContentEnabled"
+          @update:model-value="updateSpecialContentProtection"
+        />
+        <span class="preprocess-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>
+        </span>
+      </label>
    </div>
  </div>

@@ -224,36 +216,6 @@ defineExpose({ revealValidation })
        </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">
@@ -286,21 +248,13 @@ defineExpose({ revealValidation })
        </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>
+    <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>
  </div>

  <div class="form-section generation-options-section">
@@ -313,25 +267,11 @@ defineExpose({ revealValidation })
    <div class="generation-option-list">
      <GenerationOptionsPanel
        :options="options"
-        :models="generationModels"
        section="quality"
-        :validation-message="validationAttempted ? generationValidationMessage : ''"
+        :validation-message="validationAttempted ? qualityValidationMessage : ''"
        @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">
@@ -358,23 +298,6 @@ defineExpose({ revealValidation })
    </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">
@@ -419,11 +342,60 @@ defineExpose({ revealValidation })
  margin-top: 16px;
 }

-.compact-option-list {
-  gap: 8px;
+.preprocess-option-grid {
+  display: grid;
+  grid-template-columns: repeat(2, minmax(0, 1fr));
+  gap: 12px;
+  margin-top: 16px;
+}

-  .generation-option-row {
-    min-height: 58px;
+.preprocess-option {
+  display: flex;
+  gap: 10px;
+  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 {
+  display: grid;
+  gap: 4px;
+
+  strong {
+    color: #344054;
+    font-size: 13px;
+    font-weight: 600;
+  }
+
+  small {
+    color: #8a93a3;
+    font-size: 12px;
+    line-height: 1.5;
  }
 }

@@ -518,75 +490,8 @@ defineExpose({ revealValidation })
  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 {
+  .chunk-settings-grid {
    grid-template-columns: minmax(0, 1fr);
  }

@@ -595,10 +500,13 @@ defineExpose({ revealValidation })
    flex-direction: column;
  }

-  .compact-option-list .generation-option-row,
-  .advanced-protection-row {
+  .compact-option-list .generation-option-row {
    align-items: center;
    flex-direction: row;
  }
+
+  .preprocess-option-grid {
+    grid-template-columns: minmax(0, 1fr);
+  }
 }
 </style>
--- a/frontend/src/views/data-process/create/data-process-create.scss
+++ b/frontend/src/views/data-process/create/data-process-create.scss
@@ -37,7 +37,7 @@
  display: flex;
  align-items: center;
  justify-content: space-between;
-  max-width: 860px;
+  max-width: 1040px;
  margin: 0 auto;
  padding: 0 20px;
 }
@@ -45,11 +45,7 @@
 .step-item {
  display: flex;
  align-items: center;
-  flex: 1;
-
-  &:first-child {
-    flex: 0;
-  }
+  flex: none;
 }

 .step-connector {
@@ -60,14 +56,14 @@
  transition: background-color 0.3s;
 }

-.step-item.is-completed .step-connector,
-.step-item.is-active .step-connector {
+.step-connector.is-active {
  background-color: #5146e5;
 }

 .step-node {
  display: flex;
  align-items: center;
+  flex-shrink: 0;
  gap: 10px;
 }

@@ -149,6 +145,30 @@
  min-width: 160px;
 }

+@media (max-width: 1100px) {
+  .custom-wizard-steps {
+    padding: 0;
+  }
+
+  .step-connector {
+    margin: 0 8px;
+  }
+
+  .step-node {
+    gap: 7px;
+  }
+
+  .step-title {
+    display: none;
+  }
+
+  .step-item.is-active .step-title {
+    display: block;
+    font-size: 12px;
+    white-space: nowrap;
+  }
+}
+
@media (max-width: 760px) {
  .wizard-main {
    padding: 24px 20px;
@@ -168,9 +188,7 @@

  .step-title {
    max-width: 72px;
-    font-size: 12px;
    line-height: 1.35;
-    white-space: normal;
  }

  .wizard-footer {
--- a/frontend/src/views/data-process/create/types.ts
+++ b/frontend/src/views/data-process/create/types.ts
@@ -1,6 +1,6 @@
 export type ProcessType = 'structured' | 'unstructured' | 'external'

-export type StepId = 'create' | 'upload' | 'preview' | 'generate' | 'results'
+export type StepId = 'create' | 'model' | 'upload' | 'preview' | 'generate' | 'results'

 export type PreprocessOption =
  | 'clean_invalid'
--- a/frontend/src/views/data-process/create/useDataProcessDraft.ts
+++ b/frontend/src/views/data-process/create/useDataProcessDraft.ts
@@ -94,9 +94,9 @@ export function useDataProcessDraft(bindings: DraftBindings) {
      bindings.goToStep('create')
      bindings.task.name = snapshot.task?.name || ''
      bindings.task.description = snapshot.task?.description || ''
-      bindings.processType.value = snapshot.processType === 'structured' || snapshot.processType === 'external'
+      bindings.processType.value = snapshot.processType === 'unstructured' || snapshot.processType === 'external'
        ? snapshot.processType
-        : 'unstructured'
+        : 'structured'

      if (snapshot.structuredOptions) {
        bindings.structuredOptions.value = {
--- a/frontend/src/views/eval/EvalCreateView.vue
+++ b/frontend/src/views/eval/EvalCreateView.vue
@@ -5,7 +5,9 @@ import { ElMessage } from 'element-plus'
 import PageCard from '@/components/PageCard.vue'
 import EvalTaskSetupStep, { type EvalTaskSetupDraft } from './create/EvalTaskSetupStep.vue'
 import EvalRuleSetupStep, { type EvalRuleSetupDraft } from './create/EvalRuleSetupStep.vue'
-import { createDimension, getDimensionList, startEval } from '@/api/modules/eval'
+import BasicMetricSetupStep, { type BasicMetricSetupDraft } from './create/BasicMetricSetupStep.vue'
+import StartEvalStep from './create/StartEvalStep.vue'
+import { createDimension, startEval } from '@/api/modules/eval'
 import { getTrainedModels, getModelList } from '@/api/modules/model'
 import { getDatasetList } from '@/api/modules/dataset'
 import { getSystemInfo } from '@/api/modules/system'
@@ -19,15 +21,17 @@ const submitting = ref(false)
 const currentStep = ref(0)
 const taskStepRef = ref<StepExposed>()
 const ruleStepRef = ref<StepExposed>()
+const basicMetricStepRef = ref<StepExposed>()

 const WIZARD_STEPS = [
  { title: '任务配置', description: '选择模型、算力与评测数据' },
-  { title: '评测规则', description: '选择或创建评测维度' },
+  { title: '大模型评测指标', description: '配置评测模型、方式与标准' },
+  { title: '基础评测指标', description: '选择 BLEU、ROUGE 等参考指标' },
+  { title: '开始评测', description: '确认配置并启动评测任务' },
 ] as const

 const trainedModels = ref<TrainedModel[]>([])
 const evalDatasets = ref<DatasetItem[]>([])
-const dimensions = ref<Dimension[]>([])
 const evalModels = ref<ModelItem[]>([])
 const gpus = ref<GpuInfo[]>([])
 const createdDimensionId = ref<string | number>('')
@@ -42,9 +46,6 @@ const taskForm = ref<EvalTaskSetupDraft>({
 })

 const ruleForm = ref<EvalRuleSetupDraft>({
-  ruleMode: 'new',
-  dimension_id: '',
-  newDimension: {
  type: '',
  description: '',
  eval_model: '',
@@ -57,11 +58,19 @@ const ruleForm = ref<EvalRuleSetupDraft>({
  score_min: 0,
  score_max: 5,
  pass_threshold: 3,
-  },
+})
+
+const basicMetricForm = ref<BasicMetricSetupDraft>({
+  bleu_enabled: false,
+  bleu_n: 4,
+  rouge_enabled: false,
+  rouge_methods: ['rouge_1', 'rouge_2'],
+  cosine_enabled: false,
+  output_precision: 3,
 })

 watch(
-  () => ruleForm.value.newDimension,
+  ruleForm,
  () => {
    createdDimensionId.value = ''
  },
@@ -72,20 +81,18 @@ async function loadData() {
  loading.value = true
  const results = await Promise.allSettled([
    getTrainedModels(),
-    getDimensionList(),
    getDatasetList(),
    getSystemInfo(),
    getModelList(),
  ])

  if (results[0].status === 'fulfilled') trainedModels.value = results[0].value?.models || []
-  if (results[1].status === 'fulfilled') dimensions.value = results[1].value || []
-  if (results[2].status === 'fulfilled') {
-    evalDatasets.value = (results[2].value || []).filter((dataset) => dataset.type === 'eval')
+  if (results[1].status === 'fulfilled') {
+    evalDatasets.value = (results[1].value || []).filter((dataset) => dataset.type === 'eval')
  }
-  if (results[3].status === 'fulfilled') gpus.value = results[3].value?.gpu || []
-  if (results[4].status === 'fulfilled') {
-    evalModels.value = (results[4].value || []).filter((model) => model.purpose === 'evaluation')
+  if (results[2].status === 'fulfilled') gpus.value = results[2].value?.gpu || []
+  if (results[3].status === 'fulfilled') {
+    evalModels.value = (results[3].value || []).filter((model) => model.purpose === 'evaluation')
  }

  const failedCount = results.filter((result) => result.status === 'rejected').length
@@ -95,25 +102,19 @@ async function loadData() {
  loading.value = false
 }

-
-
 function buildDimensionPayload() {
-  const draft = ruleForm.value.newDimension
+  const draft = ruleForm.value
  const payload: Partial<Dimension> = {
    name: `Custom_Dim_${Date.now()}`,
    type: draft.type,
    description: draft.description,
-    eval_model: draft.type === 'text_similarity' ? undefined : draft.eval_model,
+    eval_model: draft.eval_model,
    eval_method: draft.eval_method,
-    eval_prompt: draft.type === 'text_similarity' ? undefined : draft.eval_prompt,
+    eval_prompt: draft.eval_prompt,
    is_active: draft.is_active,
    is_default: draft.is_default,
    create_time: new Date().toISOString(),
  }
-  if (draft.type === 'text_similarity') {
-    payload.bleu_n = draft.bleu_n
-    payload.output_precision = draft.output_precision
-  }
  if (draft.type === 'metric') {
    payload.score_min = draft.score_min
    payload.score_max = draft.score_max
@@ -123,8 +124,6 @@ function buildDimensionPayload() {
 }

 async function resolveDimensionId() {
-  if (ruleForm.value.ruleMode === 'baseline') return ''
-  if (ruleForm.value.ruleMode === 'existing') return ruleForm.value.dimension_id
  if (createdDimensionId.value !== '') return createdDimensionId.value

  const created = await createDimension(buildDimensionPayload())
@@ -137,24 +136,32 @@ async function resolveDimensionId() {

 async function handleSubmit() {
  if (loading.value || submitting.value) return
-  const ruleValid = await ruleStepRef.value?.validate()
-  if (!ruleValid) {
-    ElMessage.warning('请检查并完善评测规则')
-    return
-  }
-
  submitting.value = true
  try {
    const dimensionId = await resolveDimensionId()
    await startEval({
      eval_task_name: taskForm.value.eval_task_name,
-      eval_type: ruleForm.value.ruleMode === 'baseline' ? 'baseline' : 'custom',
+      eval_type: 'custom',
      model_id: taskForm.value.model_id,
      gpu_id: taskForm.value.gpu_id,
      dataset_id: taskForm.value.data_source === 'dataset' ? taskForm.value.dataset_id : '',
      dimension_id: dimensionId,
      data_source: taskForm.value.data_source,
      leaderboard: taskForm.value.leaderboard,
+      basic_metrics: {
+        bleu: {
+          enabled: basicMetricForm.value.bleu_enabled,
+          ngram: basicMetricForm.value.bleu_n,
+        },
+        rouge: {
+          enabled: basicMetricForm.value.rouge_enabled,
+          methods: basicMetricForm.value.rouge_methods,
+        },
+        cosine: {
+          enabled: basicMetricForm.value.cosine_enabled,
+        },
+        output_precision: basicMetricForm.value.output_precision,
+      },
    })
    ElMessage.success('评测任务已创建并启动')
    router.push('/model-eval')
@@ -168,17 +175,33 @@ async function handleSubmit() {

 async function handleNext() {
  if (loading.value || submitting.value) return
+  if (currentStep.value === 0) {
    const taskValid = await taskStepRef.value?.validate()
    if (!taskValid) {
      ElMessage.warning('请检查并完善任务配置')
      return
    }
-  currentStep.value = 1
+  }
+  if (currentStep.value === 1) {
+    const ruleValid = await ruleStepRef.value?.validate()
+    if (!ruleValid) {
+      ElMessage.warning('请检查并完善大模型评测指标')
+      return
+    }
+  }
+  if (currentStep.value === 2) {
+    const basicMetricValid = await basicMetricStepRef.value?.validate()
+    if (!basicMetricValid) {
+      ElMessage.warning('请检查并完善基础评测指标')
+      return
+    }
+  }
+  currentStep.value = Math.min(currentStep.value + 1, WIZARD_STEPS.length - 1)
 }

 function handleBack() {
  if (submitting.value) return
-  currentStep.value = 0
+  currentStep.value = Math.max(currentStep.value - 1, 0)
 }

 function handleCancel() {
@@ -189,22 +212,29 @@ onMounted(loadData)
 </script>

 <template>
-  <PageCard title="新建评测任务" subtitle="完成任务配置与评测规则设置后启动评测">
+  <PageCard
+    title="新建评测任务"
+    subtitle="依次配置任务、大模型评测指标与基础评测指标，确认后开始评测"
+  >
    <div class="create-wizard-layout">
      <main v-loading="loading" class="wizard-main">
        <div class="wizard-steps-container" aria-label="评测任务创建步骤">
          <div class="custom-wizard-steps">
+            <template v-for="(step, index) in WIZARD_STEPS" :key="step.title">
+              <div
+                v-if="index !== 0"
+                class="step-connector"
+                :class="{ 'is-active': currentStep >= index }"
+              ></div>
              <div
-              v-for="(step, index) in WIZARD_STEPS"
-              :key="step.title"
                class="step-item"
                :class="{
                  'is-active': currentStep === index,
                  'is-completed': currentStep > index,
                }"
                :aria-current="currentStep === index ? 'step' : undefined"
+                :aria-label="`${index + 1}. ${step.title}：${step.description}`"
              >
-              <div v-if="index !== 0" class="step-connector"></div>
                <div class="step-node">
                  <div class="step-icon" aria-hidden="true">
                    <i v-if="currentStep > index" class="fa fa-check" />
@@ -216,6 +246,7 @@ onMounted(loadData)
                  </div>
                </div>
              </div>
+            </template>
          </div>
        </div>

@@ -231,37 +262,53 @@ onMounted(loadData)
            :disabled="submitting"
          />
          <EvalRuleSetupStep
-          v-else
+            v-else-if="currentStep === 1"
            ref="ruleStepRef"
            v-model="ruleForm"
-          :dimensions="dimensions"
            :eval-models="evalModels"
-          :loading="loading"
            :disabled="submitting"
          />
+          <BasicMetricSetupStep
+            v-else-if="currentStep === 2"
+            ref="basicMetricStepRef"
+            v-model="basicMetricForm"
+            :disabled="submitting"
+          />
+          <StartEvalStep
+            v-else
+            :task="taskForm"
+            :llm-metric="ruleForm"
+            :basic-metrics="basicMetricForm"
+            :trained-models="trainedModels"
+            :eval-datasets="evalDatasets"
+            :eval-models="evalModels"
+            :gpus="gpus"
+          />
        </div>
      </main>

      <footer class="wizard-footer">
        <el-button
-          v-if="currentStep === 1"
+          v-if="currentStep > 0"
          class="footer-back"
          :disabled="submitting"
          @click="handleBack"
        >
-          <i class="fa fa-arrow-left footer-button-icon" />返回：任务配置
+          <i class="fa fa-arrow-left footer-button-icon" />返回：{{ WIZARD_STEPS[currentStep - 1].title }}
        </el-button>
        <el-button v-else class="footer-back" :disabled="submitting" @click="handleCancel">
          取消
        </el-button>

        <el-button
-          v-if="currentStep === 0"
+          v-if="currentStep < WIZARD_STEPS.length - 1"
          type="primary"
          :disabled="loading || submitting"
          @click="handleNext"
        >
-          下一步：评测规则<i class="fa fa-arrow-right footer-button-icon is-right" />
+          下一步：{{ WIZARD_STEPS[currentStep + 1].title }}<i
+            class="fa fa-arrow-right footer-button-icon is-right"
+          />
        </el-button>
        <el-button
          v-else
@@ -270,7 +317,7 @@ onMounted(loadData)
          :disabled="loading || submitting"
          @click="handleSubmit"
        >
-          创建并启动评测
+          开始评测
        </el-button>
      </footer>
    </div>
@@ -301,31 +348,26 @@ onMounted(loadData)
  display: flex;
  align-items: center;
  justify-content: space-between;
-  max-width: 760px;
+  max-width: 1120px;
  margin: 0 auto;
  padding: 0 20px;
 }

 .step-item {
  display: flex;
-  flex: 1;
+  flex: none;
  align-items: center;
 }

-.step-item:first-child {
-  flex: 0;
-}
-
 .step-connector {
  flex: 1;
  height: 2px;
-  margin: 0 20px;
+  margin: 0 12px;
  background: #e2e8f0;
  transition: background-color 0.2s ease;
 }

-.step-item.is-active .step-connector,
-.step-item.is-completed .step-connector {
+.step-connector.is-active {
  background: #5146e5;
 }

@@ -415,6 +457,12 @@ onMounted(loadData)
  margin-left: 6px;
 }

+@media (max-width: 1100px) {
+  .step-description {
+    display: none;
+  }
+}
+
@media (max-width: 760px) {
  .wizard-main {
    padding: 24px 20px;
@@ -425,10 +473,10 @@ onMounted(loadData)
  }

  .step-connector {
-    margin: 0 10px;
+    margin: 0 8px;
  }

-  .step-description {
+  .step-text {
    display: none;
  }

--- a/frontend/src/views/eval/create/DimensionFormFields.vue
+++ b/frontend/src/views/eval/create/DimensionFormFields.vue
@@ -21,19 +21,28 @@ export interface DimensionFormDraft {
  pass_threshold: number
 }

-withDefaults(
+const props = withDefaults(
  defineProps<{
    evalModels: ModelItem[]
    showDescription?: boolean
+    showStatusSettings?: boolean
+    allowedTypes?: DimensionType[]
  }>(),
  {
    showDescription: true,
+    showStatusSettings: true,
+    allowedTypes: () => ['classification', 'metric', 'text_similarity'],
  },
 )

 const form = defineModel<DimensionFormDraft>({ required: true })
+const SCORE_STEP = 0.5

-const typeOptions = Object.entries(DIMENSION_TYPE_MAP).map(([value, label]) => ({ value, label }))
+const typeOptions = computed(() =>
+  Object.entries(DIMENSION_TYPE_MAP)
+    .filter(([value]) => props.allowedTypes.includes(value as DimensionType))
+    .map(([value, label]) => ({ value, label })),
+)
 const currentMethods = computed(() => (form.value.type ? EVAL_METHODS[form.value.type] || [] : []))

 /**
@@ -65,6 +74,23 @@ watch(
  },
  { flush: 'sync' },
 )
+
+/** 始终保持评分区间合法，并把通过阈值收敛到当前区间内。 */
+watch(
+  () => [form.value.type, form.value.score_min, form.value.score_max] as const,
+  ([type, scoreMin, scoreMax]) => {
+    if (type !== 'metric') return
+    if (scoreMax <= scoreMin) {
+      form.value.score_max = scoreMin + SCORE_STEP
+      return
+    }
+    form.value.pass_threshold = Math.min(
+      Math.max(form.value.pass_threshold, scoreMin),
+      scoreMax,
+    )
+  },
+  { flush: 'sync' },
+)
 </script>

 <template>
@@ -134,13 +160,24 @@ watch(
  </template>

  <template v-if="form.type === 'metric'">
-    <el-form-item label="评分最小值">
-      <el-input-number v-model="form.score_min" :min="0" :step="1" style="width: 160px" />
+    <el-form-item label="评分最小值" prop="score_min">
+      <el-input-number
+        v-model="form.score_min"
+        :min="0"
+        :max="Math.max(form.score_max - SCORE_STEP, 0)"
+        :step="SCORE_STEP"
+        style="width: 160px"
+      />
    </el-form-item>
-    <el-form-item label="评分最大值">
-      <el-input-number v-model="form.score_max" :min="0" :step="1" style="width: 160px" />
+    <el-form-item label="评分最大值" prop="score_max">
+      <el-input-number
+        v-model="form.score_max"
+        :min="form.score_min + SCORE_STEP"
+        :step="SCORE_STEP"
+        style="width: 160px"
+      />
    </el-form-item>
-    <el-form-item label="通过阈值">
+    <el-form-item label="通过阈值" prop="pass_threshold">
      <div style="width: 100%">
        <el-slider
          v-model="form.pass_threshold"
@@ -154,6 +191,7 @@ watch(
    </el-form-item>
  </template>

+  <template v-if="showStatusSettings">
    <el-form-item label="启用该维度">
      <el-switch v-model="form.is_active" />
    </el-form-item>
@@ -161,6 +199,7 @@ watch(
      <el-switch v-model="form.is_default" />
    </el-form-item>
  </template>
+</template>

 <style scoped>
 .method-desc {
--- a/frontend/src/views/eval/create/EvalRuleSetupStep.vue
+++ b/frontend/src/views/eval/create/EvalRuleSetupStep.vue
@@ -1,49 +1,31 @@
 <script setup lang="ts">
-import { computed, ref } from 'vue'
+import { ref } from 'vue'
 import type { FormInstance, FormRules } from 'element-plus'
 import DimensionFormFields, { type DimensionFormDraft } from './DimensionFormFields.vue'
-import type { Dimension, ModelItem } from '@/types'
+import type { DimensionType, ModelItem } from '@/types'

-export type EvalRuleMode = 'existing' | 'new' | 'baseline'
+export type EvalRuleSetupDraft = DimensionFormDraft

-export interface EvalRuleSetupDraft {
-  ruleMode: EvalRuleMode
-  dimension_id: string | number
-  newDimension: DimensionFormDraft
-}
-
-const props = defineProps<{
-  dimensions: Dimension[]
+defineProps<{
  evalModels: ModelItem[]
-  loading?: boolean
  disabled?: boolean
 }>()

 const form = defineModel<EvalRuleSetupDraft>({ required: true })
-const existingFormRef = ref<FormInstance>()
 const newFormRef = ref<FormInstance>()
-
-const existingRules: FormRules = {
-  dimension_id: [{ required: true, message: '请选择评测维度', trigger: 'change' }],
-}
+const LLM_METRIC_TYPES: DimensionType[] = ['classification', 'metric']

 const dimensionRules: FormRules<DimensionFormDraft> = {
-  type: [{ required: true, message: '请选择指标类型', trigger: 'change' }],
-  eval_model: [{ required: true, message: '请选择大模型', trigger: 'change' }],
+  type: [{ required: true, message: '请选择大模型评测指标类型', trigger: 'change' }],
+  eval_model: [{ required: true, message: '请选择评测大模型', trigger: 'change' }],
  eval_method: [{ required: true, message: '请选择评估方式', trigger: 'change' }],
  eval_prompt: [{ required: true, message: '请填写评估 Prompt', trigger: 'blur' }],
 }

-const selectedDimension = computed(() =>
-  props.dimensions.find((dimension) => String(dimension.id) === String(form.value.dimension_id)),
-)
-
 async function validate() {
-  const target = newFormRef.value
-  if (!target) return false
+  if (!newFormRef.value) return false
  try {
-    const valid = await target.validate()
-    return !!valid
+    return !!(await newFormRef.value.validate())
  } catch {
    return false
  }
@@ -56,7 +38,7 @@ defineExpose({ validate })
  <div class="rule-step">
    <el-form
      ref="newFormRef"
-      :model="form.newDimension"
+      :model="form"
      :rules="dimensionRules"
      label-width="130px"
      class="step-form step-form-wide"
@@ -64,17 +46,17 @@ defineExpose({ validate })
      scroll-to-error
    >
      <DimensionFormFields
-        v-model="form.newDimension"
+        v-model="form"
        :eval-models="evalModels"
        :show-description="false"
+        :show-status-settings="false"
+        :allowed-types="LLM_METRIC_TYPES"
      />
    </el-form>
  </div>
 </template>

 <style scoped>
-
-
 .step-form {
  max-width: 760px;
 }
@@ -82,6 +64,4 @@ defineExpose({ validate })
 .step-form-wide {
  max-width: 920px;
 }
-
-
 </style>