---
license:
- apache-2.0
- cc-by-nc-sa-4.0
- mit
- unknown
task_categories:
- question-answering
language:
- en
pretty_name: math-vault
size_categories:
- 10K<n<100K
tags:
- math
- reasoning
- benchmark
- mathematics
configs:
- config_name: all
  data_files: data/all.jsonl
  default: true
- config_name: aime_2024
  data_files: data/aime_2024/train.jsonl
- config_name: aime_2025
  data_files: data/aime_2025/train.jsonl
- config_name: aime_2026
  data_files: data/aime_2026/train.jsonl
- config_name: aimo_validation_aime
  data_files: data/aimo_validation_aime/train.jsonl
- config_name: aimo_validation_amc
  data_files: data/aimo_validation_amc/train.jsonl
- config_name: amc_2023
  data_files: data/amc_2023/train.jsonl
- config_name: brumo_2025
  data_files: data/brumo_2025/train.jsonl
- config_name: cmimc_2025
  data_files: data/cmimc_2025/train.jsonl
- config_name: dapo_math_17k_dedup
  data_files: data/dapo_math_17k_dedup/train.jsonl
- config_name: gsm8k
  data_files: data/gsm8k/train.jsonl
- config_name: hmmt_feb_2025
  data_files: data/hmmt_feb_2025/train.jsonl
- config_name: math_500
  data_files: data/math_500/train.jsonl
- config_name: minerva_math
  data_files: data/minerva_math/train.jsonl
- config_name: olympiad_bench
  data_files: data/olympiad_bench/train.jsonl
- config_name: omni_math
  data_files: data/omni_math/train.jsonl
---

# math-vault

**One format. Clear provenance. Parser-verified. The current snapshot has 15 evaluation subsets, ready to eval.**

Most math datasets on HF come in different shapes — Parquet, JSON, tar.gz, custom splits, inconsistent answer formats. math-vault normalizes them all to a single canonical schema (`id` / `problem` / `answer` / `dataset`) and verifies every answer through a unified parser.

**This repo is the consumption layer.** For full source history, per-dataset provenance READMEs, build scripts, and derived variants, see the complete repository at [Geraldxm/math-vault](https://github.com/Geraldxm/math-vault).

| | Typical HF dataset | math-vault |
|---|---|---|
| **Format** | Per-dataset custom schema | Unified `id/problem/answer/dataset` JSONL |
| **Provenance** | Often unclear | Each dataset tracked to upstream — see GitHub |
| **Answer verification** | None | All answers parsed through Math-Verify 0.9.0 |
| **Ready to eval** | Requires custom loading scripts | Drop into [math-eval](https://github.com/Geraldxm/math-eval) directly |

## Quick start

```python
from datasets import load_dataset

# Load all 32,226 rows (`all` is the default config)
ds = load_dataset("Geraldxm/math-vault", split="train")

# Or load one subset directly by its named config
aime = load_dataset("Geraldxm/math-vault", "aime_2024", split="train")
```

Config names exactly match the machine-readable `dataset` values below. Each ID is globally unique: `aime_2024/60`, `gsm8k/0`, `dapo_math_17k_dedup/a5676144-...`.

The collection is open to future additions: add a normalized `data/<dataset>/train.jsonl` file and a matching config entry for each new subset, while retaining the same schema and provenance requirements.

## Dataset contents

Use the exact value in the `dataset` column when filtering. The current snapshot contains 15 evaluation subsets and 32,226 parser-valid evaluation rows. A further 19 Omni-MATH records are quarantined for audit and are not included here or in evaluation.

| Display name | `dataset` value | Rows |
|---|---|---:|
| DAPO-Math-17K Dedup | `dapo_math_17k_dedup` | 17,176 |
| GSM8K | `gsm8k` | 8,792 |
| Omni-MATH | `omni_math` | 4,409 |
| OlympiadBench | `olympiad_bench` | 674 |
| MATH-500 | `math_500` | 500 |
| Minerva-Math | `minerva_math` | 272 |
| AIMO Validation AIME | `aimo_validation_aime` | 90 |
| AIMO Validation AMC | `aimo_validation_amc` | 83 |
| AMC 2023 | `amc_2023` | 40 |
| CMIMC 2025 | `cmimc_2025` | 40 |
| AIME 2024 | `aime_2024` | 30 |
| AIME 2025 | `aime_2025` | 30 |
| AIME 2026 | `aime_2026` | 30 |
| BrUMO 2025 | `brumo_2025` | 30 |
| HMMT February 2025 | `hmmt_feb_2025` | 30 |
| **Total** | — | **32,226** |

## Licensing

The repository code and documentation are MIT-licensed. Dataset rows retain their upstream terms; this collection does not relicense them. The card metadata lists every license class present: Apache-2.0, CC BY-NC-SA 4.0, MIT, and `unknown` where an upstream source declares no license. Check the [per-dataset provenance and license records](https://github.com/Geraldxm/math-vault#%E6%95%B0%E6%8D%AE%E7%9B%AE%E5%BD%95) before use or redistribution.

## Citation

**math-vault** (this dataset):

```bibtex
@dataset{ge_math_vault_2026,
  author  = {Ge, Xinmu},
  title   = {math-vault: Curated, Traceable Snapshots of Public Mathematical Reasoning Datasets},
  year    = {2026},
  version = {v0.1.0},
  doi     = {10.5281/zenodo.21411214},
  url     = {https://doi.org/10.5281/zenodo.21411214}
}
```

**math-eval** (the evaluation framework):

```bibtex
@software{ge_math_eval_2026,
  author  = {Ge, Xinmu},
  title   = {math-eval: Reproducible Mathematical Reasoning Generation and Evaluation},
  year    = {2026},
  doi     = {10.5281/zenodo.21411207},
  url     = {https://github.com/Geraldxm/math-eval},
  license = {Apache-2.0}
}
```