mehuldamani/hotpot_qa

This dataset is a modified version of the HotPotQA distractor dataset, which contains factual questions requiring multi-hop reasoning. In the original HotPotQA dataset, each example presents ten paragraphs, only two of which contain the information necessary to answer the question; the remaining eight paragraphs include closely related but irrelevant details. Consequently, solving this task requires the model to identify and reason over the pertinent passages. To more strongly develop uncertainty reasoning capability, we construct HotPotQA-Modified, in which we systematically remove either 0, 1, or both of the key paragraphs required to answer each question. This modification introduces varying levels of informational completeness that the model must reason over. The "gold_removed" column indicates the number of relevant paragraphs removed (0,1 or 2). Questions are distributed across three equal groups: one-third have no relevant paragraphs (0/8), one-third have 1 relevant paragraph (1/7), and one-third have both relevant paragraphs (2/6). Each question consistently contains 8 total paragraphs.

To cite the original HotPotQA dataset:

@article{yang2018hotpotqa,
  title={HotpotQA: A dataset for diverse, explainable multi-hop question answering},
  author={Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W and Salakhutdinov, Ruslan and Manning, Christopher D},
  journal={arXiv preprint arXiv:1809.09600},
  year={2018}
}

To cite our modified dataset:

@article{damani2025beyond,
  title={Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty},
  author={Damani, Mehul and Puri, Isha and Slocum, Stewart and Shenfeld, Idan and Choshen, Leshem and Kim, Yoon and Andreas, Jacob},
  journal={arXiv preprint arXiv:2507.16806},
  year={2025}
}