feat: 新增外部数据源拉取与 DPO 输出格式支持

- 支持从 PostgreSQL 数据库拉取结构化数据作为训练来源
- 新增 DPO (Direct Preference Optimization) 输出类型
- 支持 chosen/rejected 字段的编辑、校验和发布
- 完善数据预处理切分逻辑和元数据管理
- 移除 OCR 扫描 PDF 功能，保持基础文本解析能力

Co-Authored-By: Claude <noreply@anthropic.com>

backend/app/modules/data_process/generation.py

@@ -29,7 +29,12 @@ class _TerminalModelGenerationError(ModelGenerationError):
 
 OUTPUT_TYPE_STANDARD = "standard"
 OUTPUT_TYPE_REASONING = "reasoning"
-SUPPORTED_OUTPUT_TYPES = {OUTPUT_TYPE_STANDARD, OUTPUT_TYPE_REASONING}
+OUTPUT_TYPE_DPO = "dpo"
+SUPPORTED_OUTPUT_TYPES = {
+    OUTPUT_TYPE_STANDARD,
+    OUTPUT_TYPE_REASONING,
+    OUTPUT_TYPE_DPO,
+}
 REASONING_DETAIL_NORMAL = "normal"
 REASONING_DETAIL_DETAILED = "detailed"
 SUPPORTED_REASONING_DETAILS = {
@@ -285,6 +290,18 @@ def _prompt_messages(
             "这是思维链输出模式，即使其他提示语要求省略分析，也不得省略 reasoning。"
             "不要自行添加 <think> 标签，系统会在保存时统一组装。"
         )
+    elif output_type == OUTPUT_TYPE_DPO:
+        schema = (
+            '{"items":[{"instruction":"...","input":"...",'
+            '"chosen":"...","rejected":"..."}]}'
+        )
+        output_rule = (
+            "你正在生成用于直接偏好优化（DPO）的成对偏好数据。"
+            "instruction、chosen 和 rejected 均不得为空；chosen 必须是忠于来源、"
+            "准确完整的优选回答，rejected 必须是表面合理但存在明确质量缺陷的拒选回答。"
+            "两者不得相同；rejected 不得包含违法危险内容，也不得用空白、乱码或无关文本凑数。"
+            "不要输出分析过程或 <think> 标签。"
+        )
     else:
         schema = '{"items":[{"instruction":"...","input":"...","output":"..."}]}'
         output_rule = (
@@ -461,9 +478,13 @@ def generate_model_records(
                             "instruction": failure_instruction,
                             "input": content,
                             "output": "",
+                            "chosen": "",
+                            "rejected": "",
                             "original_instruction": failure_instruction,
                             "original_input": content,
                             "original_output": "",
+                            "original_chosen": "",
+                            "original_rejected": "",
                             "status": "invalid",
                             "error": error_message,
                             "split": "train",
@@ -479,6 +500,8 @@ def generate_model_records(
                     input_text = normalize_text(
                         str(value.get("input") or value.get("context") or "")
                     )
+                    chosen = ""
+                    rejected = ""
                     if output_type == OUTPUT_TYPE_REASONING:
                         reasoning = normalize_text(
                             re.sub(
@@ -508,6 +531,36 @@ def generate_model_records(
                         )
                         valid = bool(instruction and reasoning and answer)
                         missing_error = "model result is missing instruction, reasoning or answer"
+                    elif output_type == OUTPUT_TYPE_DPO:
+                        chosen = normalize_text(str(value.get("chosen") or ""))
+                        rejected = normalize_text(str(value.get("rejected") or ""))
+                        chosen = normalize_text(
+                            re.sub(
+                                r"<think>[\s\S]*?(?:</think>|$)",
+                                "",
+                                chosen,
+                                flags=re.IGNORECASE,
+                            )
+                        )
+                        rejected = normalize_text(
+                            re.sub(
+                                r"<think>[\s\S]*?(?:</think>|$)",
+                                "",
+                                rejected,
+                                flags=re.IGNORECASE,
+                            )
+                        )
+                        output = chosen
+                        valid = bool(
+                            instruction
+                            and chosen
+                            and rejected
+                            and chosen.strip() != rejected.strip()
+                        )
+                        missing_error = (
+                            "model result is missing instruction, chosen or rejected, "
+                            "or chosen equals rejected"
+                        )
                     else:
                         output = normalize_text(
                             str(
@@ -527,7 +580,10 @@ def generate_model_records(
                         )
                         valid = bool(instruction and output)
                         missing_error = "model result is missing instruction or output"
-                    raw_id = f"{preview_id}:{variant_index + 1}:{instruction}:{output}"
+                    raw_id = (
+                        f"{preview_id}:{variant_index + 1}:{instruction}:"
+                        f"{output}:{rejected}"
+                    )
                     result_id = f"result_{hashlib.sha256(raw_id.encode()).hexdigest()[:16]}"
                     results.append(
                         {
@@ -536,9 +592,13 @@ def generate_model_records(
                             "instruction": instruction,
                             "input": input_text,
                             "output": output,
+                            "chosen": chosen,
+                            "rejected": rejected,
                             "original_instruction": instruction,
                             "original_input": input_text,
                             "original_output": output,
+                            "original_chosen": chosen,
+                            "original_rejected": rejected,
                             "status": "valid" if valid else "invalid",
                             "error": (None if valid else missing_error),
                             "split": "train",