MedHorizon: 340 Full-Procedure Clinical Videos for Sparse Evidence Retrieval

MedHorizon is a long-context medical video benchmark created by DBD123 and last updated on 2026-05-07. It contains 340 full-procedure clinical videos paired with 1,253 multiple-choice question-answer pairs. The benchmark is designed to evaluate multimodal models on tasks requiring sparse evidence retrieval and multi-hop reasoning across long videos.

Use Cases

• Benchmarking multimodal models on long-context video understanding based on full-procedure clinical videos.
• Evaluating sparse evidence retrieval capabilities based on the benchmark's emphasis on locating key moments in long videos.
• Testing multi-hop reasoning over distributed observations based on questions requiring reasoning across different parts of a procedure.

Strengths

• Contains 340 full-procedure clinical videos, providing long-context data.
• Includes 1,253 multiple-choice QA pairs for structured evaluation.
• Specifically designed to test sparse evidence retrieval and multi-hop reasoning, addressing gaps in short-clip datasets.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset size are unknown, which may limit suitability assessment.
• The dataset is provided as a test-only evaluation split, which may limit its use for training.

Provenance

Source

DBD123 on Hugging Face

Freshness

Last updated 2026-05-07 09:04:55; freshness should be verified.