Julian2002/RLVR-Math-16k

---
language:
- en
task_categories:
- text-generation
tags:
- math
- reasoning
- rlhf
- rlvr
- grpo
size_categories:
- 10K<n<100K
---

# RLVR-Math-16k

A curated math reasoning dataset for **RLVR (Reinforcement Learning with Verifiable Rewards)** training.

## Dataset Summary

| Split | Samples |
|-------|--------:|
| train | 16,384 |
| test | 842 |
| **Total** | **17,226** |

## Source Datasets

### train

| Source | Samples |
|--------|--------:|
| hiyouga/math12k | 10,476 |
| nlile/NuminaMath-1.5-RL-Verifiable/amc_aime | 3,075 |
| nlile/NuminaMath-1.5-RL-Verifiable/olympiads | 2,833 |

### test

| Source | Samples |
|--------|--------:|
| hiyouga/math12k | 500 |
| math-ai/minervamath | 272 |
| math-ai/amc23 | 40 |
| math-ai/aime25 | 30 |

### Training Sources

- [hiyouga/math12k](https://huggingface.co/datasets/hiyouga/math12k): MATH competition problems (converted from OpenAI PRM800K)
- [nlile/NuminaMath-1.5-RL-Verifiable](https://huggingface.co/datasets/nlile/NuminaMath-1.5-RL-Verifiable): AMC/AIME and Olympiad competition problems

### Test Sources

- [hiyouga/math12k](https://huggingface.co/datasets/hiyouga/math12k): MATH500
- [math-ai/minervamath](https://huggingface.co/datasets/math-ai/minervamath): Minerva Math
- [math-ai/aime25](https://huggingface.co/datasets/math-ai/aime25): AIME 2025
- [math-ai/amc23](https://huggingface.co/datasets/math-ai/amc23): AMC 2023

## Data Format

Each sample follows the verl-compatible chat format:

```json
{
"data_source": "source_dataset_id",
"prompt": [
{"role": "system", "content": "..."},
{"role": "user", "content": "math problem text"}
],
"ability": "math",
"reward_model": {"style": "rule", "ground_truth": "answer"},
"extra_info": {"split": "train/test", "index": 0}
}
```

## Preprocessing

**Training data filters:**
- Source filter: only competition-level problems (olympiads, amc_aime)
- Length filter: problem <= 2000 chars, solution <= 3000 chars
- Test set deduplication: removed overlapping problems with all test benchmarks
- Stratified sampling by source category
- Answer parsability: verified via [math-verify](https://github.com/huggingface/Math-Verify) to ensure reliable reward signals

**Test data:** standard benchmarks used as-is (no filtering applied).

## Intended Use

This dataset is designed for RLVR math reasoning training (e.g., DAPO, REINFORCE++) with rule-based reward verification.