ExpVid: A Benchmark for Multimodal LLMs on 390 Scientific Experiment Videos

ExpVid is a benchmark dataset for evaluating Multimodal Large Language Models on scientific experiment videos. It comprises 390 lab experiment videos spanning 13 disciplines and includes 10 tasks across 3 levels of understanding. The dataset was created by OpenGVLab and was last updated on October 14, 2025.

Use Cases

• Benchmarking MLLM performance on scientific video question-answering based on the 10 defined tasks.
• Training models for temporal reasoning in lab procedures based on the video collection.
• Evaluating cross-disciplinary video understanding based on the 13 scientific disciplines covered.
• Developing models for fine-grained action recognition in experiment videos based on the multi-level task structure.

Strengths

• Contains 390 curated lab experiment videos.
• Covers 13 distinct scientific disciplines.
• Structured into 10 benchmark tasks across 3 levels of complexity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the specific video collection.

Provenance

Source

OpenGVLab

Collection Method

Curated from a collection of lab experiment videos.

Time Range

null

Freshness

Last updated 2025-10-14 11:36:21; freshness should be verified.

Geography

null