Three Multi-Hop Question Answering Datasets: CommonsenseQA, OpenBookQA, and MedQA-USMLE

12,102 question-answer pairs in CommonsenseQA require real-world commonsense reasoning. 5,957 four-option multiple-choice questions in OpenBookQA are based on student-level scientific knowledge, and 12,723 questions from the USMLE comprise the MedQA-USMLE biomedical dataset. The collection, uploaded by daiyi Li to figshare under CC-BY-4.0, is packaged in a 466.7 MB RAR file.

Use Cases

• Benchmarking multi-step reasoning models based on the described multi-hop question answering task.
• Training models for commonsense inference based on questions derived from the ConceptNet knowledge base.
• Evaluating biomedical knowledge and clinical reasoning based on questions from the United States Medical Licensing Examination.
• Developing models that integrate external knowledge for scientific question answering based on the OpenBookQA description.

Strengths

• Includes three established benchmark datasets with defined sizes: CommonsenseQA (12,102 pairs), OpenBookQA (5,957 questions), and MedQA-USMLE (12,723 questions).
• Covers distinct reasoning domains: commonsense, scientific knowledge, and professional biomedical knowledge.
• Released under a permissive CC-BY-4.0 license, facilitating reuse and redistribution.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the combined collection is unknown, which may limit suitability assessment.
• The description focuses on the datasets' purpose for evaluating a specific model rather than detailed data schema.

Provenance

Source

Aggregated from three public datasets: CommonsenseQA, OpenBookQA, and MedQA-USMLE.

Collection Method

Likely compiled for research evaluation, as described in the input.

Freshness

Last updated 2026-04-30 05:30:16; freshness should be verified.