VSTAT5S: 500 Spatiotemporal Reasoning Questions on 450 Synthetic 5-Second Videos

A 2026 dataset by ShushengYang contains 500 question-answer pairs for evaluating multimodal AI models. It is a short-video companion to VSTAT, featuring 450 synthetic video clips trimmed to approximately 5 seconds each. The dataset is packaged for use with the lmms-eval framework.

Use Cases

• Benchmarking video understanding models based on 500 synthetic video QA pairs.
• Evaluating spatiotemporal reasoning capabilities based on tasks rendered in Blender.
• Testing model performance on short-form video content based on clips trimmed to ~5 seconds.
• Assessing numerical and multiple-choice question answering based on 200 numeric and 300 multiple-choice questions.

Strengths

• Contains 500 total QA pairs, with a defined split of 200 numeric and 300 multiple-choice questions.
• Includes 450 synthetic video clips, each approximately 5 seconds long.
• Packaged for plug-and-play use with the lmms-eval framework, following established layout conventions.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the video files is unknown, which may limit suitability assessment.
• Data is synthetic, which may limit generalizability to real-world video understanding tasks.

Provenance

Source

ShushengYang on Hugging Face.

Collection Method

Synthetic tasks rendered using Blender.

Freshness

Last updated 2026-06-10 18:31:25.