# License Attribution for Eval Datasets

This directory contains subsets of three public datasets used for deterministic graph evaluation gates. All datasets are licensed under permissive open-source licenses.

## MuSiQue-Ans

**Source:** https://github.com/askplatypus/musique

**License:** Creative Commons Attribution 4.0 International (CC BY 4.0)

**Citation:**
```
@article{trivedi2021musique,
  title={MuSiQue: Multihop Questions via Single-hop Question Composition},
  author={Trivedi, Harsh and Bisk, Yonatan and Schwartz, Roy and Bhatnagar, Sameer and Khot, Tushar},
  journal={arXiv preprint arXiv:2108.00573},
  year={2021}
}
```

**Subset:** ~100 two-hop questions with supporting paragraphs, deterministically selected via fixed-seed stratified sampling.

---

## 2WikiMultiHopQA

**Source:** https://github.com/Alibaba-NLP/2wikimultihopqa

**License:** Apache License 2.0 (Apache-2.0)

**Citation:**
```
@article{ho2020constructing,
  title={Constructing Datasets for Multi-hop Reading Comprehension Across Documents},
  author={Ho, Vinh and Nguyen, Khoa D and Cao, Siu Cheung and Lukasiewicz, Thomas},
  journal={Transactions of the Association for Computational Linguistics},
  volume={8},
  pages={287--303},
  year={2020},
  publisher={MIT Press}
}
```

**Subset:** ~100 multi-hop questions (~50 bridge, ~50 comparison) with supporting paragraphs, deterministically selected via fixed-seed stratified sampling.

---

## HotpotQA

**Source:** https://hotpotqa.github.io/

**License:** Creative Commons Attribution 4.0 International (CC BY 4.0)

**Citation:**
```
@inproceedings{yang2018hotpotqa,
  title={HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering},
  author={Yang, Zhilin and Qi, Priyanka and Manning, Christopher D and Salakhutdinov, Ruslan},
  booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  pages={2369--2380},
  year={2018}
}
```

**Subset:** ~100 distractor-type questions with supporting paragraphs, deterministically selected via fixed-seed sampling.

---

## Summary

| Dataset | Collection Name | Records | License |
|---------|-----------------|---------|---------|
| MuSiQue | `multihop-musique` | ~100 | CC BY 4.0 |
| 2WikiMultiHopQA | `multihop-2wiki` | ~100 | Apache-2.0 |
| HotpotQA | `hotpotqa` | ~100 | CC BY 4.0 |

All subsets are committed to the repository as plain JSONL files under `tests/eval/corpus/` with no Git LFS required. The fixtures are used exclusively by the deterministic graph evaluation gates in `tests/eval/test_e2e_graph_eval_gate_v2.py` and are not shipped in production.
