language:
- en
tags:
- vision-language-action
- gaming-agent
- fps
- action-chunk
- event-driven-control
license: other
GamePlayer V16 Broad EventChunk-SFT
This is the frozen V16 broad multi-game SFT mixture used in the EventChunk-FPS research lineage. It is an archival training release, not the later V31 causal unified-development set and not a claim of current benchmark results.
Contents
- Train: 1,351,283 conversation rows, 10,677,608 effective assistant turns.
- Validation: 73,980 conversation rows, 586,786 effective assistant turns.
- Total: 1,425,263 rows and 11,264,394 image/action supervision turns.
- 23,509 train and 1,359 validation trajectory groups, with zero reported train/validation trajectory overlap and zero missing image paths.
- Action horizons: 8, 12, 16, or 24 atomic actions. The environment is intended to pause during generation and each chunk executes once.
The data comes from OpenP2P, CrossFPS, CS2, Gaming500, and a small ViZDoom
success-anchor set. The mix is intentionally broad but imbalanced: most rows
come from OpenP2P. See sft_v16_scale100_mixture.manifest.json and
validation.json for exact source and game counts.
Layout
The JSONL files are under:
game/research/formal/artifacts/v16_scales/sft_v16_scale100_mixture/
Their image references are repository-relative paths rooted at
game/research/formal/. The complete release is stored under archives/ as
uncompressed tar shards plus SHA256SUMS, avoiding millions of separate Hub
objects. Verify the checksums and extract every shard from the directory where
the dataset checkout lives, for example: for f in archives/*.tar; do tar -xf "$f"; done. This recreates the original directory layout so the images
values resolve without rewriting.
Format
Rows use ShareGPT-style multi-turn conversations. images contains one RGB
path per decision; each assistant turn contains a strict action-chunk DSL:
<p>m=... h=... i=...</p>
<a>k=... x=... y=... l=... r=... m=...</a>
...
<eoc>
Provenance and Use
This release contains gameplay media derived from multiple source datasets and video games. Downstream users must follow the terms of every source dataset and the applicable game publisher policies. In particular, this card does not grant ownership of game assets or waive third-party rights. Preserve upstream attribution when redistributing derived material.
This historical V16 label contract predates the stricter V31 causal-data gate; it should not be mixed with V31 results or represented as a V31 training set.