HotpotQA: Synthetic Long-Context Multi-Hop Question Answering Data

HotpotQA data is enriched with synthetic long-context examples to push the boundaries of multi-hop reasoning. The dataset, provided by BytedTsinghua-SIA, is designed for training and evaluating long-context language models like MemAgent. It was last updated on July 30, 2025.

Use Cases

• Training long-context LLMs based on synthetic multi-hop QA data
• Evaluating model performance on main QA tasks
• Testing model robustness on out-of-distribution (OOD) tasks
• Benchmarking multi-hop reasoning capabilities in extended contexts

Strengths

• Derived from the established HotpotQA dataset, providing a known foundation
• Specifically designed for long-context LLM evaluation, targeting a current research challenge
• Includes out-of-distribution tasks for robustness testing

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

BytedTsinghua-SIA

Collection Method

Primarily derived from HotpotQA and enriched with synthetic long-context data

Time Range

null

Freshness

Last updated 2025-07-30 06:57:24

Geography

null