VPT Owamcap: Minecraft Video and Action Data for Agent Training

Open-world-agents provides the VPT Owamcap dataset, a conversion of the Video PreTraining (VPT) Minecraft dataset into the OWAMcap format. It contains 26,322 samples with a total volume of 5.2 terabytes, each pairing a video file with corresponding action logs. The dataset was last updated on July 10, 2025.

Use Cases

• Training video-based imitation learning agents based on the paired .mp4 video and action data.
• Developing or benchmarking multimodal AI models that process visual input and sequential action output.
• Researching agent behavior in simulated 3D environments like Minecraft using the provided action sequences.
• Converting or working with action data in the OWAMcap format mentioned in the description.

Strengths

• Contains 26,322 individual samples, providing a substantial base for model training.
• Offers a large data volume of 5.2 terabytes, suggesting high-resolution or lengthy video content.
• Each sample provides multimodal alignment with video (.mp4) and corresponding actions in two formats (.jsonl, .mcap).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality and sample specifics require manual inspection after download.

Provenance

Source

Converted from the original Video PreTraining (VPT) Minecraft dataset.

Collection Method

Conversion to OWAMcap format and compression with WebDataset Format.

Time Range

null

Freshness

Last updated 2025-07-10 04:59:30; freshness should be verified.

Geography

null