VLM Eval Videos: 693 Short Video Clips for Action Recognition Benchmarking

VLM Eval Videos is a benchmark dataset containing 693 short MP4 video clips for evaluating Vision-Language Models. The dataset, created by author gnitoahc, is organized into five categories, with each clip paired with a fixed question and a ground-truth short-sentence answer. It was last updated on the Hugging Face platform in June 2026.

Use Cases

• Benchmarking VLM performance on short-form action recognition based on the paired video-question-answer structure.
• Automating VLM inference scoring using the provided ground-truth answers and an LLM-as-a-judge setup.
• Evaluating model generalization across the five distinct video categories mentioned in the description.

Strengths

• Contains 693 video clips, providing a defined scale for evaluation.
• Each clip is paired with a fixed question and ground-truth answer, enabling standardized automated scoring.
• Clips are organized into five categories, suggesting a structured evaluation across different action types.

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

Hugging Face (user gnitoahc).

Collection Method

Clips are drawn from YouTube.

Freshness

Last updated 2026-06-04 06:00:34; freshness should be verified.