VideoRLVR-Data: Procedurally Generated Visual Reasoning Tasks

VideoRLVR-Data provides training and evaluation data for verifiable video reasoning. The dataset contains procedurally generated visual reasoning tasks from three domains: Maze, FlowFree, and Sokoban. It was created by DarthZhu and last updated on HuggingFace in May 2026.

Use Cases

• Training video models to generate rule-consistent visual trajectories based on the described reasoning tasks.
• Evaluating model performance on automatically verifiable visual reasoning tasks.
• Benchmarking video models across the Maze, FlowFree, and Sokoban domains mentioned in the description.
• Developing reinforcement learning environments with verifiable rewards for video generation.

Strengths

• Data is built for a specific, stricter research question regarding rule-consistency and automatic verification.
• Contains tasks from three distinct procedural generation domains: Maze, FlowFree, and Sokoban.
• Last updated on May 20, 2026, indicating recent maintenance.

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

HuggingFace dataset repository created by DarthZhu.

Collection Method

Procedurally generated.

Freshness

Last updated 2026-05-20 06:10:32