jankin123 commited on
Commit
9dbedc4
·
verified ·
1 Parent(s): 9917889

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +2 -3
README.md CHANGED
@@ -10,7 +10,7 @@ tags:
10
  license: other
11
  ---
12
 
13
- # GamePlayer Broad EventChunk-SFT
14
 
15
  This is the frozen broad multi-game SFT mixture used in the EventChunk-FPS
16
  research lineage.
@@ -22,8 +22,7 @@ research lineage.
22
  - Total: 1,425,263 rows and 11,264,394 image/action supervision turns.
23
  - 23,509 train and 1,359 validation trajectory groups, with zero reported
24
  train/validation trajectory overlap and zero missing image paths.
25
- - Action horizons: 8, 12, 16, or 24 atomic actions. The environment is
26
- intended to pause during generation and each chunk executes once.
27
 
28
  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
29
  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.
 
10
  license: other
11
  ---
12
 
13
+ # GamePlayer-1.4M
14
 
15
  This is the frozen broad multi-game SFT mixture used in the EventChunk-FPS
16
  research lineage.
 
22
  - Total: 1,425,263 rows and 11,264,394 image/action supervision turns.
23
  - 23,509 train and 1,359 validation trajectory groups, with zero reported
24
  train/validation trajectory overlap and zero missing image paths.
25
+ - Action horizons: 8, 12, 16, or 24 atomic actions.
 
26
 
27
  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
28
  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.