src/sequence_data_prepare.py

import sys
import os
import json
import argparse
from datasets import load_dataset
from prompt_pool import DATASET_PROMPTS, SEQUENCE_PROMPT



def self_critique_prompt(question:str, response:str):
    prompt = SEQUENCE_PROMPT["self_critique_new"]
    prompt = prompt.format(input_data=question, previous_solution=response)
    return prompt

def concate_prompt(question:str, response:list[str]):
    for i, r in enumerate(response):
        response[i] = f"Solution {i+1}: {r}"
    response = "\n".join(response)
    prompt = SEQUENCE_PROMPT["concate"]
    prompt = prompt.format(input_data=question, previous_solution=response)
    return prompt


def prepare_sequence_jsonl(input_path, output_dir, budget=8, model_name="Qwen/Qwen2.5-7B-Instruct", max_completion_tokens=2048):
    # input_path: jsonl file
    # output_dir: directory to save the jsonl file
    # model_name: model name
    # max_completion_tokens: max completion tokens
    # 预留出batch空间，也即为后续data jsonl batch数量
    
    # check if output_dir exists
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)
    
    batchs= [[] for _ in range(budget)]
    with open(input_path, "r", encoding="utf-8") as f:
        for idx, line in enumerate(f):
            line = json.loads(line)
            if len(line["outputs"]) < budget:
                raise ValueError(f"Warning: {line['id']} has less than {budget} outputs")
            for i in range(budget):
                item = {
                    "custom_id": f"request-{idx+1}",
                    "method": "POST",
                    "url": "/v1/chat/completions",
                    "body": {
                        "model": model_name,
                        "messages": [
                            {"role": "system", "content": "You are a helpful assistant."},
                            {"role": "user", "content": self_critique_prompt(line["problem"], line["outputs"][i])}
                        ],
                        "max_completion_tokens": max_completion_tokens,
                        "temperature": 0.7,
                    }
                }
                batchs[i].append(item)

    for i in range(budget):
        with open(os.path.join(output_dir, f"batch_{i}.jsonl"), "w", encoding="utf-8") as f:
            for line in batchs[i]:
                f.write(json.dumps(line, ensure_ascii=False) + '\n')

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_path", type=str, required=True)
    parser.add_argument("--output_dir", type=str, required=True)
    parser.add_argument("--budget", type=int, default=8)
    parser.add_argument("--model_name", type=str, default="Qwen/Qwen2.5-7B-Instruct")
    parser.add_argument("--max_completion_tokens", type=int, default=2048)
    args = parser.parse_args()
    prepare_sequence_jsonl(args.input_path, args.output_dir, args.budget, args.model_name, args.max_completion_tokens)

if __name__ == "__main__":
    main()