feat(data-process): 增加可配置思维链生成

2026-07-27 14:41:38 +08:00
parent 97cdb5cc68
commit f97245b814
12 changed files with 224 additions and 29 deletions
--- a/frontend/scripts/regression-data-process-detail.mjs
+++ b/frontend/scripts/regression-data-process-detail.mjs
@@ -88,6 +88,9 @@ assert.match(detailSource, /key !== 'generation_model_snapshot'/, '处理配置
 assert.match(detailSource, /preprocessOptionLabelMap/, '处理配置没有把预处理内部枚举转换为中文')
 assert.match(detailSource, /output_type:\s*'输出类型'/, '处理配置没有显示输出类型名称')
 assert.match(detailSource, /value === 'reasoning' \? '思维链回答' : '标准回答'/, '处理配置没有转换输出类型枚举')
+assert.match(detailSource, /reasoning_detail:\s*'推理详细程度'/, '处理配置没有显示推理详细程度名称')
+assert.match(detailSource, /value === 'detailed' \? '详细推理' : '普通推理'/, '处理配置没有转换推理详细程度枚举')
+assert.match(detailSource, /key !== 'reasoning_detail' \|\| config\.output_type === 'reasoning'/, '标准回答任务不应展示无关的推理详细程度')
 assert.match(detailSource, /const parsed = typeof value === 'number' \? value : Number\(value\)/, '详情页不兼容 PostgreSQL 数字字符串')
 assert.match(detailSource, /new Date\(startTime\.value\)\.getTime\(\)/, '详情页没有在后端耗时缺失时按开始、完成时间回算')
 assert.match(detailSource, /outputCount \+ numeric\(detail\.value\?\.filtered_count\)/, '结果保留率没有使用输出和过滤结果的同口径分母')
--- a/frontend/scripts/regression-data-process-wizard.mjs
+++ b/frontend/scripts/regression-data-process-wizard.mjs
@@ -461,6 +461,7 @@ for (const field of [
  'generationModelId',
  'generationPrompt',
  'outputType',
+  'reasoningDetail',
  'qualityFilterEnabled',
  'filterLowQuality',
  'filterShortContent',
@@ -481,12 +482,14 @@ assert.match(modelSelectionSource, /class="form-section"/, '大模型选择步
 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', '输出类型', '标准回答', '思维链回答', '质量筛选', '过滤低质量内容', '过滤过短内容', '最少字数']) {
+for (const label of ['大模型', '数据生成模型', '默认提示语', '输出类型', '标准回答', '思维链回答', '推理详细程度', '普通推理（推荐）', '详细推理', '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, /aria-label="输出类型"/, '输出类型选项缺少可访问名称')
+assert.match(generationControlSource, /v-if="options\.outputType === 'reasoning'" class="output-type-row"/, '推理详细程度没有按思维链模式渐进显示')
+assert.match(generationControlSource, /aria-label="推理详细程度"/, '推理详细程度缺少可访问名称')
 assert.match(generationControlSource, /&lt;think&gt;推理过程&lt;\/think&gt;/, '思维链选项没有说明最终保存格式')
 assert.match(generationControlSource, /<el-select[\s\S]*?class="output-type-select"[\s\S]*?aria-label="输出类型"/, '输出类型必须使用右侧下拉选择')
 assert.doesNotMatch(generationControlSource, /class="output-type-options"/, '输出类型不应继续使用横向按钮组')
@@ -511,9 +514,12 @@ assert.match(
  /\.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.equal((stateSource.match(/outputType:\s*'standard'/g) || []).length, 2, '结构化与非结构化任务应默认生成标准回答')
+assert.match(stateSource, /const DEFAULT_STANDARD_GENERATION_PROMPT\s*=\s*['"][^'"]{40,}['"]/, '标准回答缺少独立默认提示语')
+assert.match(stateSource, /const DEFAULT_REASONING_GENERATION_PROMPT\s*=\s*['"][^'"]{40,}['"]/, '思维链回答缺少独立默认提示语')
+assert.equal((stateSource.match(/generationPrompt:\s*DEFAULT_STANDARD_GENERATION_PROMPT/g) || []).length, 2, '结构化与非结构化任务应默认使用标准回答提示语')
+assert.match(generationControlSource, /isBuiltInGenerationPrompt\(props\.options\.generationPrompt\)[\s\S]*?defaultGenerationPrompt\(outputType\)[\s\S]*?: props\.options\.generationPrompt/, '切换输出类型时没有在保留自定义提示语的前提下切换内置提示语')
+assert.equal((stateSource.match(/^\s{4}outputType:\s*'standard',/gm) || []).length, 2, '结构化与非结构化任务应默认生成标准回答')
+assert.equal((stateSource.match(/reasoningDetail:\s*'normal'/g) || []).length, 2, '结构化与非结构化任务应默认使用普通推理')
 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')
@@ -687,6 +693,7 @@ for (const [backendField, frontendField] of [
  ['generation_model_id', 'generationModelId'],
  ['generation_prompt', 'generationPrompt'],
  ['output_type', 'outputType'],
+  ['reasoning_detail', 'reasoningDetail'],
  ['temperature', 'temperature'],
  ['max_tokens', 'maxTokens'],
  ['json_mode', 'jsonMode'],
@@ -730,7 +737,9 @@ assert.match(stateSource, /Object\.prototype\.hasOwnProperty\.call\(config, key\
 assert.match(stateSource, /Number\.isFinite\(value\) \? value : fallback/, '配置反向映射没有保留合法数字 0')
 assert.match(stateSource, /qaPairsPerRow:\s*normalizeQaPairsGenerationCount\([\s\S]*?qa_pairs_per_row[\s\S]*?defaults\.qaPairsPerRow/, '结构化生成数量回填没有按 1 到 50 归一化')
 assert.match(stateSource, /qaPairsPerChunk:\s*normalizeQaPairsGenerationCount\([\s\S]*?qa_pairs_per_chunk[\s\S]*?defaults\.qaPairsPerChunk/, '非结构化生成数量回填没有按 1 到 50 归一化')
-assert.match(stateSource, /outputType:\s*configValue\(config, 'output_type', defaults\.outputType\) === 'reasoning'[\s\S]*?\? 'reasoning'[\s\S]*?: 'standard'/, '输出类型没有从任务配置安全回填')
+assert.match(stateSource, /const outputType = configValue\(config, 'output_type', defaults\.outputType\) === 'reasoning'[\s\S]*?\? 'reasoning'[\s\S]*?: 'standard'/, '输出类型没有从任务配置安全回填')
+assert.match(stateSource, /reasoningDetail:\s*configValue\(config, 'reasoning_detail', defaults\.reasoningDetail\) === 'detailed'[\s\S]*?\? 'detailed'[\s\S]*?: 'normal'/, '推理详细程度没有从任务配置安全回填')
+assert.match(stateSource, /isBuiltInGenerationPrompt\(configuredPrompt\)[\s\S]*?defaultGenerationPrompt\(outputType\)/, '旧版内置提示语没有按输出类型迁移')
 assert.match(stateSource, /createStructuredOptionsFromConfig/, '结构化配置缺少后端到表单的反向映射')
 assert.match(stateSource, /createUnstructuredOptionsFromConfig/, '非结构化配置缺少后端到表单的反向映射')
 assert.match(stateSource, /configValue<unknown>\(config, 'preprocess_options', \[\]\)/, '历史任务缺少预处理配置时必须按后端空列表语义回填')
@@ -808,6 +817,7 @@ for (const field of [
  'generationModelId',
  'generationPrompt',
  'outputType',
+  'reasoningDetail',
  'qualityFilterEnabled',
  'filterLowQuality',
  'filterShortContent',
--- a/frontend/src/types/dataProcess.ts
+++ b/frontend/src/types/dataProcess.ts
@@ -5,6 +5,7 @@ export type DataProcessType = 'structured' | 'unstructured' | 'external'
 export type DataProcessResultStatus = 'valid' | 'modified' | 'invalid'
 export type DataProcessSplit = 'train' | 'validation' | 'test'
 export type DataProcessOutputType = 'standard' | 'reasoning'
+export type DataProcessReasoningDetail = 'normal' | 'detailed'

 export interface DataProcessPage<T> {
  items: T[]
@@ -22,6 +23,7 @@ export interface DataProcessDatasetSplit {
 export type DataProcessConfig = Record<string, unknown> & {
  dataset_split?: DataProcessDatasetSplit
  output_type?: DataProcessOutputType
+  reasoning_detail?: DataProcessReasoningDetail
 }

 export interface DataProcessTask {
--- a/frontend/src/views/data-process/DataProcessCreateView.vue
+++ b/frontend/src/views/data-process/DataProcessCreateView.vue
@@ -202,6 +202,7 @@ function toBackendConfig(): DataProcessConfig {
    generation_model_id: options.generationModelId,
    generation_prompt: options.generationPrompt,
    output_type: options.outputType,
+    reasoning_detail: options.reasoningDetail,
    temperature: options.temperature,
    max_tokens: options.maxTokens,
    json_mode: options.jsonMode,
@@ -210,7 +211,6 @@ function toBackendConfig(): DataProcessConfig {
    filter_short_content: options.filterShortContent,
    min_output_length: options.minOutputLength,
  }
-
  if (processType.value === 'unstructured') {
    return {
      ...common,
--- a/frontend/src/views/data-process/DataProcessDetailView.vue
+++ b/frontend/src/views/data-process/DataProcessDetailView.vue
@@ -71,6 +71,7 @@ const configLabelMap: Record<string, string> = {
  generation_model_id: '数据生成模型',
  generation_prompt: '生成提示语',
  output_type: '输出类型',
+  reasoning_detail: '推理详细程度',
  temperature: '生成温度',
  max_tokens: '最大输出长度',
  json_mode: 'JSON 输出',
@@ -225,15 +226,19 @@ const durationText = computed(() => {
  return minutes ? `${minutes} 分 ${restSeconds} 秒` : `${restSeconds} 秒`
 })

-const configRows = computed(() => Object.entries(detail.value?.config || {})
-  .filter(([key]) => (
-    key !== 'generation_model_snapshot'
-    && !/(?:password|secret|token|api_key)/i.test(key)
-  ))
-  .map(([key, value]) => ({
-    label: configLabelMap[key] || key.split('_').join(' '),
-    value: formatConfigValue(key, value),
-  })))
+const configRows = computed(() => {
+  const config = detail.value?.config || {}
+  return Object.entries(config)
+    .filter(([key]) => (
+      key !== 'generation_model_snapshot'
+      && (key !== 'reasoning_detail' || config.output_type === 'reasoning')
+      && !/(?:password|secret|token|api_key)/i.test(key)
+    ))
+    .map(([key, value]) => ({
+      label: configLabelMap[key] || key.split('_').join(' '),
+      value: formatConfigValue(key, value),
+    }))
+})

 function formatConfigValue(key: string, value: unknown) {
  if (key === 'generation_model_id') {
@@ -249,6 +254,9 @@ function formatConfigValue(key: string, value: unknown) {
  if (key === 'output_type') {
    return value === 'reasoning' ? '思维链回答' : '标准回答'
  }
+  if (key === 'reasoning_detail') {
+    return value === 'detailed' ? '详细推理' : '普通推理'
+  }
  if (key === 'dataset_split' && value && typeof value === 'object') {
    const split = value as Partial<DataProcessDatasetSplit>
    return `训练集 ${split.train ?? 0}% / 验证集 ${split.validation ?? 0}% / 测试集 ${split.test ?? 0}%`
--- a/frontend/src/views/data-process/create/GenerationOptionsPanel.vue
+++ b/frontend/src/views/data-process/create/GenerationOptionsPanel.vue
@@ -1,7 +1,10 @@
 <script setup lang="ts">
-import { ref } from 'vue'
 import type { ModelItem } from '@/types'
 import type { GenerationControlOptions } from './types'
+import {
+  defaultGenerationPrompt,
+  isBuiltInGenerationPrompt,
+} from './dataProcessCreateState'

 const props = defineProps<{
  options: GenerationControlOptions
@@ -31,7 +34,18 @@ function updateQualityRules(value: Array<string | number>) {
 }

 function updateOutputType(value: string | number | boolean | undefined) {
-  updateField('outputType', value === 'reasoning' ? 'reasoning' : 'standard')
+  const outputType = value === 'reasoning' ? 'reasoning' : 'standard'
+  emit('update:options', {
+    ...props.options,
+    outputType,
+    generationPrompt: isBuiltInGenerationPrompt(props.options.generationPrompt)
+      ? defaultGenerationPrompt(outputType)
+      : props.options.generationPrompt,
+  })
+}
+
+function updateReasoningDetail(value: string | number | boolean | undefined) {
+  updateField('reasoningDetail', value === 'detailed' ? 'detailed' : 'normal')
 }

 const selectedQualityRules = () => [
@@ -91,7 +105,10 @@ function modelMeta(model: ModelItem) {
    <div class="model-field">
      <div class="field-copy">
        <strong>默认提示语</strong>
-        <small>用于约束生成内容的格式、语气和完整性，可按任务需要修改</small>
+        <small v-if="options.outputType === 'reasoning'">
+          当前使用思维链专用提示语；系统还会按所选详细程度约束推理结构
+        </small>
+        <small v-else>当前使用标准回答提示语，只要求问题和最终答案</small>
      </div>
      <el-input
        class="prompt-input"
@@ -177,9 +194,28 @@ function modelMeta(model: ModelItem) {
          <el-option label="思维链回答" value="reasoning" />
        </el-select>
      </div>
+      <div v-if="options.outputType === 'reasoning'" class="output-type-row">
+        <div class="field-copy">
+          <strong>推理详细程度</strong>
+          <small>控制推理过程保留关键步骤，或完整展开依据与中间推导</small>
+        </div>
+        <el-select
+          class="output-type-select"
+          :model-value="options.reasoningDetail"
+          aria-label="推理详细程度"
+          @update:model-value="updateReasoningDetail"
+        >
+          <el-option label="普通推理（推荐）" value="normal" />
+          <el-option label="详细推理" value="detailed" />
+        </el-select>
+      </div>
      <p class="output-type-hint">
        <template v-if="options.outputType === 'reasoning'">
-          生成结果将按 <code>&lt;think&gt;推理过程&lt;/think&gt;</code> 加最终答案的格式保存。
+          <template v-if="options.reasoningDetail === 'detailed'">
+            完整展开条件、来源依据、中间推导和结论核对，
+          </template>
+          <template v-else>保留关键依据与必要步骤，</template>
+          最终按 <code>&lt;think&gt;推理过程&lt;/think&gt;</code> 加最终答案保存。
        </template>
        <template v-else>仅保存最终答案，不包含推理过程。</template>
      </p>
--- a/frontend/src/views/data-process/create/dataProcessCreateState.ts
+++ b/frontend/src/views/data-process/create/dataProcessCreateState.ts
@@ -9,7 +9,25 @@ import type {
 } from './types'
 import { normalizeQaPairsGenerationCount } 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。仅输出符合所选输出类型和目标格式的内容，答案应事实清晰、语言自然，不要添加无关说明。'
+const LEGACY_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。仅输出符合所选输出类型和目标格式的内容，答案应事实清晰、语言自然，不要添加无关说明。'
+
+export const DEFAULT_STANDARD_GENERATION_PROMPT = '你是一名专业的数据生成助手。请严格依据输入内容生成准确、完整、可直接用于监督微调的问答数据。只生成问题和最终答案，不输出分析、推理过程或来源中不存在的信息；答案应事实清晰、语言自然。'
+
+export const DEFAULT_REASONING_GENERATION_PROMPT = '你是一名专业的推理数据生成助手。请严格依据输入内容生成问题、可核验的推理过程和最终答案。推理需要说明关键依据与必要步骤，不得引入来源中不存在的事实；最终答案应准确、完整且语言自然。'
+
+export function defaultGenerationPrompt(outputType: 'standard' | 'reasoning') {
+  return outputType === 'reasoning'
+    ? DEFAULT_REASONING_GENERATION_PROMPT
+    : DEFAULT_STANDARD_GENERATION_PROMPT
+}
+
+export function isBuiltInGenerationPrompt(value: string) {
+  return [
+    LEGACY_DEFAULT_GENERATION_PROMPT,
+    DEFAULT_STANDARD_GENERATION_PROMPT,
+    DEFAULT_REASONING_GENERATION_PROMPT,
+  ].includes(value)
+}

 export function createDefaultStructuredOptions(): StructuredProcessOptions {
  return {
@@ -18,8 +36,9 @@ export function createDefaultStructuredOptions(): StructuredProcessOptions {
    qaPairsPerRow: 1,
    datasetSplit: { train: 80, validation: 10, test: 10 },
    generationModelId: '',
-    generationPrompt: DEFAULT_GENERATION_PROMPT,
+    generationPrompt: DEFAULT_STANDARD_GENERATION_PROMPT,
    outputType: 'standard',
+    reasoningDetail: 'normal',
    temperature: 0.7,
    maxTokens: 1024,
    jsonMode: false,
@@ -52,8 +71,9 @@ export function createDefaultUnstructuredOptions(): UnstructuredProcessOptions {
    qaPairsPerChunk: 1,
    datasetSplit: { train: 80, validation: 10, test: 10 },
    generationModelId: '',
-    generationPrompt: DEFAULT_GENERATION_PROMPT,
+    generationPrompt: DEFAULT_STANDARD_GENERATION_PROMPT,
    outputType: 'standard',
+    reasoningDetail: 'normal',
    temperature: 0.7,
    maxTokens: 1024,
    jsonMode: false,
@@ -92,12 +112,21 @@ function generationOptionsFromConfig(
  config: DataProcessConfig,
  defaults: GenerationControlOptions,
 ): GenerationControlOptions {
+  const outputType = configValue(config, 'output_type', defaults.outputType) === 'reasoning'
+    ? 'reasoning'
+    : 'standard'
+  const configuredPrompt = String(
+    configValue(config, 'generation_prompt', defaults.generationPrompt),
+  )
  return {
    generationModelId: configValue(config, 'generation_model_id', defaults.generationModelId),
-    generationPrompt: String(configValue(config, 'generation_prompt', defaults.generationPrompt)),
-    outputType: configValue(config, 'output_type', defaults.outputType) === 'reasoning'
-      ? 'reasoning'
-      : 'standard',
+    generationPrompt: isBuiltInGenerationPrompt(configuredPrompt)
+      ? defaultGenerationPrompt(outputType)
+      : configuredPrompt,
+    outputType,
+    reasoningDetail: configValue(config, 'reasoning_detail', defaults.reasoningDetail) === 'detailed'
+      ? 'detailed'
+      : 'normal',
    temperature: numberValue(config, 'temperature', defaults.temperature),
    maxTokens: numberValue(config, 'max_tokens', defaults.maxTokens),
    jsonMode: Boolean(configValue(config, 'json_mode', defaults.jsonMode)),
@@ -224,6 +253,7 @@ export function generationAffectingOptionsFor(
    generationModelId: options.generationModelId,
    generationPrompt: options.generationPrompt,
    outputType: options.outputType,
+    reasoningDetail: options.reasoningDetail,
    temperature: options.temperature,
    maxTokens: options.maxTokens,
    jsonMode: options.jsonMode,
--- a/frontend/src/views/data-process/create/types.ts
+++ b/frontend/src/views/data-process/create/types.ts
@@ -1,4 +1,8 @@
-import type { DataProcessOutputType, DataProcessPreviewFileStatus } from '@/types/dataProcess'
+import type {
+  DataProcessOutputType,
+  DataProcessPreviewFileStatus,
+  DataProcessReasoningDetail,
+} from '@/types/dataProcess'

 export type ProcessType = 'structured' | 'unstructured' | 'external'

@@ -33,6 +37,7 @@ export interface GenerationControlOptions {
  generationModelId: string | number | ''
  generationPrompt: string
  outputType: DataProcessOutputType
+  reasoningDetail: DataProcessReasoningDetail
  temperature: number
  maxTokens: number
  jsonMode: boolean