#!/usr/bin/env python3 """Prepare GSM8K and MATH parquet files in VERL's standard RL format. Mirrors VERL's examples/data_preprocess/{gsm8k,math_dataset}.py so that the MI250X runs use the same prompt templates and rule rewards as the stock VERL examples. """ from __future__ import annotations import argparse import re from pathlib import Path import datasets import pyarrow.parquet as pq GSM8K_INSTRUCTION = 'Let\'s think step by step and output the final answer after "####".' MATH_INSTRUCTION = "Let's think step by step and output the final answer within \\boxed{}." def last_boxed_only_string(text: str) -> str | None: index = text.rfind("\\boxed") if index < 0: index = text.rfind("\\fbox") if index < 0: return None depth = 0 for position in range(index, len(text)): if text[position] == "{": depth += 1 elif text[position] == "}": depth -= 1 if depth == 0: return text[index:position + 1] return None def remove_boxed(text: str) -> str: if text.startswith("\\boxed "): return text[len("\\boxed "):] left = "\\boxed{" assert text.startswith(left) and text.endswith("}"), text return text[len(left):-1] def write(data: datasets.Dataset, path: Path) -> None: # Drop the Hugging Face feature metadata: older `datasets` releases in # the training container cannot parse feature types written by newer ones. path.parent.mkdir(parents=True, exist_ok=True) pq.write_table(data.data.table.replace_schema_metadata(None), str(path)) def gsm8k(output: Path) -> None: source = "openai/gsm8k" loaded = datasets.load_dataset(source, "main") def convert(split): def row(example, index): answer = re.search(r"#### (\-?[0-9\.\,]+)", example["answer"]).group(1).replace(",", "") return { "data_source": source, "prompt": [{"role": "user", "content": example["question"] + " " + GSM8K_INSTRUCTION}], "ability": "math", "reward_model": {"style": "rule", "ground_truth": answer}, "extra_info": {"split": split, "index": index}, } return row for split, name in (("train", "train"), ("test", "test")): data = loaded[split].map(convert(split), with_indices=True, remove_columns=loaded[split].column_names) write(data.flatten_indices(), output / "gsm8k" / f"{name}.parquet") def math(output: Path) -> None: source = "DigitalLearningGmbH/MATH-lighteval" loaded = datasets.load_dataset(source, trust_remote_code=True) def convert(split): def row(example, index): return { "data_source": source, "prompt": [{"role": "user", "content": example["problem"] + " " + MATH_INSTRUCTION}], "ability": "math", "reward_model": {"style": "rule", "ground_truth": remove_boxed(last_boxed_only_string(example["solution"]))}, "extra_info": {"split": split, "index": index}, } return row for split in ("train", "test"): # A few solutions carry no boxed answer and cannot be rule-scored. boxed = loaded[split].filter(lambda example: last_boxed_only_string(example["solution"]) is not None) data = boxed.map(convert(split), with_indices=True, remove_columns=boxed.column_names) write(data.flatten_indices(), output / "math" / f"{split}.parquet") def main() -> None: parser = argparse.ArgumentParser() parser.add_argument("--output", type=Path, required=True) args = parser.parse_args() gsm8k(args.output) math(args.output) if __name__ == "__main__": main()