File size: 1,479 Bytes
849b59c
 
 
 
 
 
 
 
 
 
 
 
9dbedc4
849b59c
9917889
 
849b59c
 
 
 
 
 
 
 
9dbedc4
849b59c
9917889
 
849b59c
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
---
language:
- en
tags:
- vision-language-action
- gaming-agent
- fps
- action-chunk
- event-driven-control
license: other
---

# GamePlayer-1.4M

This is the frozen broad multi-game SFT mixture used in the EventChunk-FPS
research lineage.

## 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 source data comes from [OpenP2P](https://huggingface.co/elefantai/open-p2p), [CrossFPS](https://huggingface.co/datasets/zizhaotong/CrossFPS-train), [CS2](https://huggingface.co/datasets/RekaAI/CS2-10k), [Gaming500](https://huggingface.co/datasets/markov-ai/gaming-500-hours), and a small ViZDoom
success-anchor set we created. You can get the source video from these repo, cause we just provide the frames from codec. See `sft_v16_scale100_mixture.manifest.json` and `validation.json` under `archives/metadata.tar` for exact source and game counts.

## Format

Rows use ShareGPT-style multi-turn conversations. `images` contains one RGB
path per decision; each assistant turn contains a strict action-chunk DSL:

```text
<p>m=... h=... i=...</p>
<a>k=... x=... y=... l=... r=... m=...</a>
...
<eoc>
```