make_grpo_dataset.py

import argparse
import os

import datasets
from datasets import load_dataset
from verl.utils.hdfs_io import copy, makedirs
from verl.utils.reward_score.math import last_boxed_only_string, remove_boxed


def extract_solution(solution_str):
    return remove_boxed(last_boxed_only_string(solution_str))


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--local_dir", default="./data/math")

    args = parser.parse_args()

    # 'lighteval/MATH' is no longer available on huggingface.
    # Use mirror repo: DigitalLearningGmbH/MATH-lighteval
    data_source = "DigitalLearningGmbH/MATH-lighteval"  
    train_dataset = load_dataset("json", data_files='grpo_1.5B_without_sub_data.json', split="train[100:]")
    test_dataset = load_dataset("json", data_files='grpo_1.5B_without_sub_data.json', split="train[:100]")
    instruction_following = "Let's think step by step and output the final answer within \\boxed{}."

    # add a row to each data item that represents a unique id
    def make_map_fn(split):
        def process_fn(example, idx):
            question = example.pop("question")

            question = question + " " + instruction_following
            solution = example['gold']    
            if example['type']=='sub':
                solution = eval(solution)
            else:
                solution = [solution]
            data = {
                "data_source": data_source,
                "prompt": [{"role": "user", "content": question}],
                "ability": "math",
                "reward_model": {"style": "rule", "ground_truth": solution},
                "extra_info": {"split": split, "index": idx},
            }
            return data

        return process_fn

    train_dataset = train_dataset.map(function=make_map_fn("train"), with_indices=True)
    test_dataset = test_dataset.map(function=make_map_fn("test"), with_indices=True)

    local_dir = args.local_dir

    train_dataset.to_parquet(os.path.join(local_dir, "grpo_mid_train.parquet"))
    test_dataset.to_parquet(os.path.join(local_dir, "grpo_mid_test.parquet"))
    for i in range(5):
        print(train_dataset[i])
    print(len(train_dataset))
    print(len(test_dataset))