# Data licenses

The code in `data/build/` is licensed under Apache-2.0, like the rest of the repository's code. The prompt sets in
`data/student/` and `data/teacher/` mix rows from public datasets, which keep their original licenses, with synthetic
rows that we release. Licenses were read from the Hugging Face dataset cards on 2026-09-28.

| Rows | Source | License |
|---|---|---|
| math: every `domain == "math"` row | [zwhe99/DeepMath-103K](https://huggingface.co/datasets/zwhe99/DeepMath-103K) (rev. `5cf055d`) | MIT |
| code: every `domain == "code"` row | [PRIME-RL/Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data) (rev. `9776b13`) | MIT |
| IF: every `domain == "ifeval"` row | generated by `data/build/gen_ifeval_data.py` (ours) | CC BY 4.0 |
| `data/build/ids/*.json`, `data/MANIFEST.json` | ours | CC BY 4.0 |

## Code sub-sources

Eurus-2-RL-Data takes its code problems from four datasets. The `metadata.data_source` field of every code row names
its sub-source.

| `data_source` | Dataset named by Eurus-2-RL-Data | License on its dataset card |
|---|---|---|
| `codecontests` | [deepmind/code_contests](https://huggingface.co/datasets/deepmind/code_contests) | CC BY 4.0 |
| `taco` | [BAAI/TACO](https://huggingface.co/datasets/BAAI/TACO) | Apache-2.0 |
| `apps` | [codeparrot/apps](https://huggingface.co/datasets/codeparrot/apps) | MIT |
| `codeforces` | [MatrixStudio/Codeforces-Python-Submissions](https://huggingface.co/datasets/MatrixStudio/Codeforces-Python-Submissions) | unverified: the card states no license |

The problem statements in these datasets come from competitive-programming sites, including Codeforces, AtCoder,
CodeChef, Kattis and others. Their copyright stays with the original authors. We redistribute them only as part of
the upstream datasets, under the terms those datasets state.

## Intermediate copy

The paper's pools were read from [Keven16/G-OPD-Training-Data](https://huggingface.co/datasets/Keven16/G-OPD-Training-Data)
(rev. `c9bc678`). Its card states Apache-2.0.

Its math and code files add nothing beyond the operations described in `data/README.md`:
- filtering and deduplicating DeepMath-103K;
- dropping the Eurus system message;
- appending a one-line instruction to each prompt.

`build_data.py --source upstream` therefore rebuilds every row from the two MIT-licensed originals alone.

## Attribution

If you use these prompt sets, please cite the upstream datasets:
- DeepMath-103K (He et al., 2025);
- Eurus-2-RL-Data / PRIME (Cui et al., 2025);
- CodeContests (Li et al., 2022);
- TACO (Li et al., 2023);
- APPS (Hendrycks et al., 2021).

Please also cite the DN-MOPD paper for the IF prompts and the selection.