VR-MultiStep: ~360k Programmatic Training Samples for Multi-Step Image-to-Video Reasoning

Approximately 360,000 programmatically generated samples form a training corpus for long-horizon multi-step image-to-video reasoning. The dataset, created by Mark7121983123, is a companion to a separate 180-instance evaluation benchmark and was last updated on May 7, 2026. It covers 36 parameterized tasks across reasoning families like Navigation, Planning, CSP, Execution, Geometry, and Physics.

Use Cases

• Training models for long-horizon image-to-video generation based on the described multi-step reasoning tasks.
• Benchmarking AI performance on complex visual planning and execution problems across the six reasoning families mentioned.
• Developing and evaluating parameterized task-solving agents using the 36 distinct task types described.
• Researching programmatic data generation methods for creating large-scale, structured reasoning corpora.

Strengths

• Large scale with approximately 360,000 total samples.
• Structured into 36 distinct parameterized tasks, providing organized coverage.
• Covers multiple reasoning families (Navigation, Planning, CSP, Execution, Geometry, Physics), suggesting breadth in problem types.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality and sample structure require manual inspection after download.
• The total size of ~164 GB is significant and may require substantial storage and bandwidth.

Provenance

Source

huggingface

Collection Method

Programmatically generated training corpus.

Freshness

Last updated 2026-05-07 00:01:44; freshness should be verified.