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>
This commit is contained in:
caoxiaozhu
2026-08-11 14:17:45 +08:00
parent f809825a7d
commit 5f6e7523cf
26 changed files with 927 additions and 144 deletions

View File

@@ -2,6 +2,7 @@ from __future__ import annotations
from enum import StrEnum
from typing import Any, Literal
from urllib.parse import parse_qs, urlsplit
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
@@ -15,6 +16,31 @@ def _config_value(config: dict[str, Any], snake_name: str, camel_name: str, defa
def _validate_process_config(config: dict[str, Any]) -> None:
output_type = _config_value(config, "output_type", "outputType", "standard")
if output_type not in {"standard", "reasoning", "dpo"}:
raise ValueError("output_type must be one of: standard, reasoning, dpo")
source_mode = _config_value(config, "source_mode", "sourceMode", "local")
if source_mode not in {"local", "external"}:
raise ValueError("source_mode must be one of: local, external")
external_source = _config_value(config, "external_source", "externalSource", None)
if external_source is not None:
if not isinstance(external_source, dict):
raise ValueError("external_source must be an object")
if any(
key.lower() in {"password", "secret", "token", "api_key"}
for key in external_source
):
raise ValueError("external_source must not persist credentials")
external_url = str(external_source.get("url") or "").strip()
if external_url:
parsed_external_url = urlsplit(external_url)
sensitive_query_keys = {"password", "secret", "token", "api_key", "user", "username"}
if parsed_external_url.username or parsed_external_url.password or (
set(parse_qs(parsed_external_url.query)) & sensitive_query_keys
):
raise ValueError("external_source URL must not contain credentials")
chunk_method = _config_value(config, "chunk_method", "chunkMethod", "layout_hybrid")
if not isinstance(chunk_method, str) or chunk_method not in {
"layout_hybrid",
@@ -303,6 +329,9 @@ class ExternalSourceRequest(BaseModel):
username: str | None = Field(default=None, max_length=150)
password: str | None = Field(default=None, max_length=500)
limit: int = Field(default=1000, ge=1, le=100_000)
connect_timeout_seconds: int = Field(default=5, ge=1, le=30)
statement_timeout_seconds: int = Field(default=30, ge=1, le=300)
ssl_mode: Literal["disable", "prefer", "require", "verify-ca", "verify-full"] = "prefer"
class ExternalPullRequest(ExternalSourceRequest):
@@ -324,6 +353,8 @@ class ResultUpdate(BaseModel):
instruction: str | None = None
input: str | None = None
output: str | None = None
chosen: str | None = None
rejected: str | None = None
expected_updated_at: str | None = None
@@ -374,7 +405,7 @@ class PublishRequest(BaseModel):
dataset_type: Literal["train", "test", "eval", "val", "other"] = "train"
storage_type: Literal["local"] = "local"
split: DatasetSplit = Field(default_factory=DatasetSplit)
format: Literal["alpaca_jsonl", "jsonl"] = "alpaca_jsonl"
format: Literal["alpaca_jsonl", "jsonl", "dpo"] = "alpaca_jsonl"
description: str = ""
@field_validator("dataset_name")