asr_rl_prepare_data.py

#!/usr/bin/env python3
"""
步骤 2：将 v4_lite 训练数据转成 verl GRPO 所需的 parquet 格式。

用法：
    python asr_rl_prepare_data.py \
        --input  /mnt/.../asr_v4_lite_20260308_164454/train.jsonl \
        --output /mnt/.../asr_rl_train.parquet \
        --max_samples 15000 \
        --max_errors 3          # 只保留 1~3 个错误的样本，太难的 RL 也学不了
"""

import argparse
import json
import re
from pathlib import Path

import pandas as pd


PROMPT_PREFIX = "你是一个ASR诗词纠错专家，纠正语音识别输出中的错误，并输出正确的诗词，输入句子为："


def count_errors(source: str, target: str) -> int:
    """粗略统计字符差异数。"""
    return sum(a != b for a, b in zip(source, target)) + abs(len(source) - len(target))


def convert(input_path: str, output_path: str, max_samples: int, max_errors: int):
    records = []
    skipped_neg = 0
    skipped_hard = 0

    with open(input_path, "r", encoding="utf-8") as f:
        for line in f:
            d = json.loads(line)
            human_val = d["conversations"][0]["value"]
            gpt_val   = d["conversations"][1]["value"]

            source = human_val.replace(PROMPT_PREFIX, "", 1).strip()
            target = gpt_val.strip()

            # 跳过负例（source == target，RL 不需要）
            if source == target:
                skipped_neg += 1
                continue

            # 跳过错误数过多的样本
            if count_errors(source, target) > max_errors:
                skipped_hard += 1
                continue

            records.append({
                "data_source": "asr_poetry",
                "prompt": [
                    {"role": "user", "content": human_val}
                ],
                "reward_model": {
                    "ground_truth": target,
                    "style": "asr_correction",
                },
                "extra_info": {
                    "source": source,
                    "target": target,
                },
            })

            if len(records) >= max_samples:
                break

    print(f"Total records: {len(records)}")
    print(f"Skipped negatives: {skipped_neg}")
    print(f"Skipped too-hard (>{max_errors} errors): {skipped_hard}")

    Path(output_path).parent.mkdir(parents=True, exist_ok=True)
    df = pd.DataFrame(records)
    df.to_parquet(output_path, index=False)
    print(f"Saved to: {output_path}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--input",
        default="./asr/log1/asr_v4_lite_20260308_164454/train.jsonl")
    parser.add_argument("--output",
        default="./asr/check/asr_rl_data/asr_rl_train.parquet")
    parser.add_argument("--val_input",
        default="./asr/log1/asr_v4_lite_20260308_164454/test_real_asr.jsonl")
    parser.add_argument("--val_output",
        default="./asr/check/asr_rl_data/asr_rl_val.parquet")
    parser.add_argument("--max_samples", type=int, default=3000)
    parser.add_argument("--max_errors",  type=int, default=3)
    args = parser.parse_args()

    print("=== Converting train set ===")
    convert(args.input, args.output, args.max_samples, args.max_errors)

    print("\n=== Converting val set ===")
    convert(args.val_input, args.val_output, max_samples=10000, max_errors=99)