#!/usr/bin/env python3 import argparse import hashlib import json import os import re from typing import Any, Dict, Iterable, List, Optional import pandas as pd from datasets import load_dataset import math import numpy as np # ----------------------------- # 工具函数 # ----------------------------- def to_jsonable(obj: Any) -> Any: """ 递归把 numpy / pandas 对象转成 json 可序列化的原生 Python 类型。 """ # None / 基本类型 if obj is None or isinstance(obj, (str, int, bool)): return obj if isinstance(obj, float): if math.isnan(obj) or math.isinf(obj): return None return obj # numpy 标量 if isinstance(obj, np.generic): value = obj.item() if isinstance(value, float) and (math.isnan(value) or math.isinf(value)): return None return value # numpy 数组 if isinstance(obj, np.ndarray): return [to_jsonable(x) for x in obj.tolist()] # pandas 缺失值 / numpy 缺失值 try: if pd.isna(obj): return None except Exception: pass # dict if isinstance(obj, dict): return {str(k): to_jsonable(v) for k, v in obj.items()} # list / tuple / set if isinstance(obj, (list, tuple, set)): return [to_jsonable(x) for x in obj] # 兜底 return str(obj) def normalize_text(x: Any) -> str: if x is None: return "" s = str(x) s = s.replace("\r\n", "\n").replace("\r", "\n") s = re.sub(r"\s+", " ", s).strip() return s def safe_dict(x: Any) -> Dict[str, Any]: return x if isinstance(x, dict) else {} def safe_list(x: Any) -> List[Any]: return x if isinstance(x, list) else [] def extract_prompt_text(prompt_obj: Any) -> str: """ 将 prompt(list[dict]) 抽取为纯文本。 例如： [{"role": "user", "content": "..."}] -> "..." """ prompt_list = safe_list(prompt_obj) parts: List[str] = [] for item in prompt_list: if isinstance(item, dict): content = item.get("content") if content is not None: parts.append(str(content)) elif item is not None: parts.append(str(item)) return "\n\n".join(p.strip() for p in parts if str(p).strip()) def make_dedup_key(row: Dict[str, Any]) -> str: """ 优先用原始题面 raw_problem 去重；若缺失，则退化到 prompt_text； 再不行则对整行做稳定哈希。 """ raw_problem = normalize_text(row.get("problem")) if raw_problem: return f"problem::{raw_problem}" prompt_text = normalize_text(row.get("prompt_text")) if prompt_text: return f"prompt::{prompt_text}" stable = json.dumps(row, ensure_ascii=False, sort_keys=True) return "hash::" + hashlib.sha256(stable.encode("utf-8")).hexdigest() def ensure_parent_dir(path: str) -> None: parent = os.path.dirname(os.path.abspath(path)) if parent: os.makedirs(parent, exist_ok=True) # ----------------------------- # 数据读取 # ----------------------------- def load_from_hf(repo_id: str, split: str) -> pd.DataFrame: ds = load_dataset(repo_id, split=split) # 数据量很小（页面显示 960 rows），这里直接转 DataFrame 即可。 return ds.to_pandas() # ----------------------------- # 数据整理与去重 # ----------------------------- def flatten_records(df: pd.DataFrame) -> pd.DataFrame: rows: List[Dict[str, Any]] = [] for _, rec in df.iterrows(): reward_model = safe_dict(rec.get("reward_model")) extra_info = safe_dict(rec.get("extra_info")) row = { "data_source": rec.get("data_source"), "ability": rec.get("ability"), "prompt_text": extract_prompt_text(rec.get("prompt")), "problem": extra_info.get("raw_problem"), "answer": reward_model.get("ground_truth"), "style": reward_model.get("style"), "source_index": extra_info.get("index"), "split": extra_info.get("split"), # 保留原始嵌套字段，方便后续追溯 "prompt": rec.get("prompt"), "reward_model": reward_model, "extra_info": extra_info, "__index_level_0__": rec.get("__index_level_0__"), } row["dedup_key"] = make_dedup_key(row) rows.append(row) out = pd.DataFrame(rows) return out def deduplicate(df: pd.DataFrame) -> pd.DataFrame: # 先按原始顺序稳定排序，保证 keep='first' 时行为可预期 sort_cols = [c for c in ["__index_level_0__", "source_index"] if c in df.columns] if sort_cols: df = df.sort_values(by=sort_cols, kind="stable").reset_index(drop=True) deduped = df.drop_duplicates(subset=["dedup_key"], keep="first").reset_index(drop=True) deduped.insert(0, "id", range(len(deduped))) return deduped # ----------------------------- # 输出 # ----------------------------- def write_jsonl(records: Iterable[Dict[str, Any]], path: str) -> None: ensure_parent_dir(path) with open(path, "w", encoding="utf-8") as f: for row in records: f.write(json.dumps(to_jsonable(row), ensure_ascii=False) + "\n") def main() -> None: parser = argparse.ArgumentParser(description="下载/读取 HF 数据集，去重后导出 JSONL 和 Parquet") parser.add_argument("--repo-id", default="BytedTsinghua-SIA/AIME-2024", help="Hugging Face dataset repo id") parser.add_argument("--split", default="train", help="数据集 split，默认 train") parser.add_argument("--save-raw-parquet", default="", help="可选：保存未去重原始 parquet") parser.add_argument("--jsonl", required=True, help="输出 jsonl 路径") parser.add_argument("--parquet", required=True, help="输出 parquet 路径") args = parser.parse_args() raw_df = load_from_hf(args.repo_id, args.split) if args.save_raw_parquet: ensure_parent_dir(args.save_raw_parquet) raw_df.to_parquet(args.save_raw_parquet, index=False) flat_df = flatten_records(raw_df) dedup_df = deduplicate(flat_df) # 导出前选择更适合作为下游训练/评测输入的列。 output_cols = [ "id", "problem", "answer", "prompt_text", "data_source", "ability", "style", "source_index", "split", "dedup_key", "prompt", "reward_model", "extra_info", "__index_level_0__", ] output_cols = [c for c in output_cols if c in dedup_df.columns] final_df = dedup_df[output_cols].copy() # JSONL records = final_df.to_dict(orient="records") write_jsonl(records, args.jsonl) # Parquet ensure_parent_dir(args.parquet) final_df.to_parquet(args.parquet, index=False) print(f"原始行数: {len(raw_df)}") print(f"去重后行数: {len(final_df)}") print(f"删除重复行数: {len(raw_df) - len(final_df)}") print(f"JSONL 已写出: {args.jsonl}") print(f"Parquet 已写出: {args.parquet}") if __name__ == "__main__": main()