---
license: mit
language:
- en
task_categories:
- text-generation
tags:
- math
- reasoning
- reinforcement-learning
- deepmath
pretty_name: DeepMath RLCSD Format Difficulty Splits
---

# DeepMath RLCSD-format difficulty splits

This dataset is derived from [`zwhe99/DeepMath-103K`](https://huggingface.co/datasets/zwhe99/DeepMath-103K)
at revision `5cf055d1fe3d7a2eb19719ac020211469736ae44` and uses the nested parquet schema released by
[`Leyiii/RLCSD`](https://huggingface.co/datasets/Leyiii/RLCSD).

## Files

| Path | Rows | Difficulty |
|---|---:|---|
| `deepmath_full/train.parquet` | 103,022 | Full upstream dataset |
| `deepmath_easy_10k/train.parquet` | 10,000 | `1.0 <= difficulty < 5.0` |
| `deepmath_medium_10k/train.parquet` | 10,000 | `5.0 <= difficulty < 7.0` |
| `deepmath_hard_10k/train.parquet` | 10,000 | `7.0 <= difficulty <= 10.0` |

The three subsets are pairwise disjoint by difficulty. Rows with missing required
fields and every occurrence of a duplicated normalized question are excluded from
the three sampling pools. Within each difficulty band, samples are selected
proportionally by exact half-step difficulty and then by topic.
Selection is deterministic (seed `42`) and uses a SHA-256
ordering over the normalized question, source index, and pinned source revision.

## Schema

Each parquet has exactly the RLCSD-style fields `prompt`, `reward_model`,
`data_source`, `ability`, and `extra_info`. `extra_info.solution` is selected uniformly from DeepMath's three R1 solutions
with Python's `random.seed(42)` behavior; `reward_model.ground_truth` comes from
`final_answer`.

## License and attribution

The upstream DeepMath-103K dataset is released under the MIT license. Please
also follow its citation and attribution requirements.