P2P: 8,300+ Hours of Annotated Gameplay from 40+ 3D Video Games

8,300+ hours of gameplay recorded at 20 FPS by experienced players across more than 40 popular 3D video games. Each frame is annotated with keyboard and mouse actions, and text instructions are provided when available. The dataset was created by elefantai and last updated on Hugging Face in March 2026.

Use Cases

• Train action-policy models based on annotated keyboard and mouse actions per frame.
• Develop vision-language-action agents based on the combination of visual frames and text instructions.
• Benchmark world model performance using sequential gameplay data recorded at 20 FPS.
• Fine-tune gaming agents based on high-quality human demonstrations from experienced players.

Strengths

• Contains over 8,300 hours of annotated data, providing a substantial training resource.
• Spans more than 40 distinct 3D video games, offering diversity in environments and tasks.
• Data is recorded at a consistent 20 FPS, providing a structured temporal sequence.
• Annotations include both low-level actions (keyboard/mouse) and high-level text instructions where available.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific model training.
• Data may reflect bias inherent to the specific 40+ games and the skill level of the experienced players who recorded it.

Provenance

Source

elefantai via Hugging Face.

Collection Method

Gameplay recorded at 20 FPS by experienced players, with frames annotated for actions and text instructions.

Freshness

Last updated 2026-03-25 00:57:56; freshness should be verified.

Geography

Region is tagged as 'us', but specific geographic coverage is not detailed.