# MATH-500 by Hugging Face H4: AI Training Dataset
Source: https://savrn.com/datasets/math-500
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Dataset Card

This dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their Let's Verify Step by Step paper. See their GitHub repo for the source file: https://github.com/openai/prm800k/tree/main?tab=readme-ov-file#math-splits

Excerpt from the card by Hugging Face H4.

## Structure

### default 500 rows

| Split | Rows | Size |
| --- | --- | --- |
| test | 500 | 400.3 KB |

problemstringsolutionstringanswerstringsubjectstringlevelint64unique_idstring

## Details

Repository

HuggingFaceH4/MATH-500

Publisher

Hugging Face H4

Task category

Text generation

Tags

Not stated by the source

Size category

Not stated by the source

Languages

en

Revision

6e4ed1a2a79af7d8630a6b768ec859cb5af4d3be

Last updated

2025-12-15

## SAVRN's Notes on MATH-500

MATH-500 is exactly what the name says: 500 problems, 450.3 KB, one config, cut from the MATH benchmark by OpenAI for the 'Let's Verify Step by Step' paper and hosted by Hugging Face H4. The inbound list shows what it is used for — math RLVR and small reasoning models (MoA-400M, MoA-150M, next-1b, next-12b-i1, and a family of Qwen2.5-3B Sheldon-RLVR checkpoints) all disclose training against it.

It is an evaluation set, and its smallness is the point: 500 problems is cheap to run at every checkpoint, which is why it shows up in so many training runs. Two cautions. The license field on our page reads 'not stated by the source', so a commercial run should clear the underlying MATH terms rather than assume. And because it is a fixed public subset of MATH test items, any model trained on MATH-family data needs an overlap check before its MATH-500 scores mean anything.

SAVRN Research

## Files

4 files, 450.3 KB in total.

Data1 file · 446.6 KB

Documentation1 file · 412 B

Other1 file · 949 B

Repository1 file · 2.4 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| test.jsonl | Data | 446.6 KB | — |
| README.md | Documentation | 412 B | — |
| eval.yaml | Other | 949 B | — |
| .gitattributes | Repository | 2.4 KB | — |

## License and Download

License

Not stated by the source

Access

No access gate

[Download from Hugging Face H4](https://huggingface.co/datasets/HuggingFaceH4/MATH-500)

Released by Hugging Face H4 through its official repository on Hugging Face.

## Models Trained on This Dataset

- Trained on (disclosed)[MoA-400M](https://savrn.com/models/moa-400m)
- Trained on (disclosed)[MoA-150M](https://savrn.com/models/moa-150m)
- Trained on (disclosed)[next-1b](https://savrn.com/models/next-1b)
- Trained on (disclosed)[next-12b-i1-GGUF](https://savrn.com/models/next-12b-i1-gguf)
- Trained on (disclosed)[Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1](https://savrn.com/models/qwen2-5-3b-instruct-sheldon-rlvr-math-v1)
- Trained on (disclosed)[Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v1-LoRA](https://savrn.com/models/qwen2-5-3b-instruct-sheldon-rlvr-math-v1-lora)
- Trained on (disclosed)[Qwen2.5-3B-Instruct-Sheldon-RLVR-math-v2-fullft](https://savrn.com/models/qwen2-5-3b-instruct-sheldon-rlvr-math-v2-fullft)

## Hugging Face H4

[All models and datasets](https://savrn.com/model-publishers/huggingfaceh4)

## Explore More

- [The dataset directory](https://savrn.com/datasets)
- [The model directory](https://savrn.com/models)
- [Research papers](https://savrn.com/papers)

## Source

- Repository metadata, read 2026-10-19.
- [Hugging Face record](https://huggingface.co/datasets/HuggingFaceH4/MATH-500)
- [How the hub is built](https://savrn.com/model-hub/methodology)