script/format_trans/raw_json2sharegpt.py

import json
import re

INPUT_PATH = "/root/test/weitiao/data_process_bq/data/parquet_all.json"
OUTPUT_PATH = "/root/test/weitiao/data_process_bq/data/sharegpt_all.json"

ROLE_MAP = {
    "system": "system",
    "user": "human",
    "assistant": "gpt"
}

def parse_prompt(prompt_text):
    """
    把 <|im_start|>xxx ... <|im_end|> 解析成 ShareGPT conversations
    """
    pattern = re.compile(
        r"<\|im_start\|>(system|user|assistant)\n(.*?)<\|im_end\|>",
        re.DOTALL
    )

    conversations = []
    for role, content in pattern.findall(prompt_text):
        conversations.append({
            "from": ROLE_MAP[role],
            "value": content.strip()
        })

    return conversations


def convert_item(item):
    # 用 chosen_prompt 解析历史对话
    conversations = parse_prompt(item["chosen_prompt"])

    # 去掉最后一个 assistant（因为真正的回复在 chosen / rejected）
    if conversations and conversations[-1]["from"] == "gpt":
        conversations = conversations[:-1]

    return {
        "conversations": conversations,
        "chosen": {
            "from": "gpt",
            "value": item["chosen"]
        },
        "rejected": {
            "from": "gpt",
            "value": item["reject"]
        }
    }


def main():
    with open(INPUT_PATH, "r", encoding="utf-8") as f:
        data = json.load(f)

    output = []
    for item in data:
        output.append(convert_item(item))

    with open(OUTPUT_PATH, "w", encoding="utf-8") as f:
        json.dump(output, f, ensure_ascii=False, indent=2)


if __name__ == "__main__":
    main()