dsa1dsa12/verl-aha-moment-dataset

Verl "Aha Moment" Dataset for DeepSeek-R1 Reproduction

Overview

This dataset is built for use with the Verl framework to reproduce the "aha moment" phenomenon observed in DeepSeek-R1-Zero. It combines two Hugging Face math-reasoning datasets into a single Parquet file conforming to the format specified in format.json.

Composition

Format

Each row in verl_deepscaler.parquet matches the schema below:

{
  "data_source": "DeepScaleR | aha-moment-750",
  "prompt": [{"role": "user", "content": "<problem text>"}],
  "ability": "math",
  "reward_model": {"style": "rule", "ground_truth": "<answer>"},
  "extra_info": {"index": 0, "solution": "<full solution text>"}
}

Key Notes

• —DeepScaleR rows: ground_truth comes directly from the dataset's answer field.
• —Aha-moment rows: ground_truth is extracted from the \boxed{...} content inside the solution (3 of 750 had no boxed answer and received an empty string).
• —All solutions are preserved verbatim in extra_info.solution for use as reference during RL training.
• —The aha-moment-750 subset is placed at indices 40,315–41,064 so it can be isolated for targeted SFT before RL.

Usage

from datasets import load_dataset

ds = load_dataset("dsa1dsa12/verl-aha-moment-dataset", split="train")

# Isolate aha-moment subset
aha_subset = ds.filter(lambda r: r["data_source"] == "aha-moment-750")

Files

• —verl_deepscaler.parquet — the combined dataset (41,065 records, ~10 MB)
• —format.json — the expected format specification
• —README.md — this file