Upload README.md with huggingface_hub

README.md

@@ -9,14 +9,17 @@ tags:
 
 # Multi-Drive World Model Data
 
-Action-conditioned driving gameplay for training a multi-game world model, grouped by visual **theme** rather than game name. 1,474,041 frames across 155 clips.
+Action-conditioned driving gameplay from multiple racing games, grouped by visual theme, for training a multi-game world model. 1,474,041 frames across 155 clips.
 
-| theme | source games | frames |
-|---|---|---|
-| cartoon | supertuxkart | 589,560 |
-| realistic | forza-horizon, need-for-speed | 585,250 |
-| arcade | asphalt-9 | 299,231 |
+| theme | source games | clips | frames |
+|---|---|---|---|
+| cartoon | supertuxkart | 118 | 589,560 |
+| realistic | forza-horizon, need-for-speed | 27 | 585,250 |
+| arcade | asphalt-9 | 10 | 299,231 |
 
-Each `data/<theme>/<clip>.tar` unpacks to `<clip>/frames/frame_*.jpg` + `<clip>/actions.jsonl` (lines: `{frame_file, actions:[accel,brake,left,right,drift], theme, game}`). `metadata.jsonl` indexes every clip.
+## Format
 
-Models trained on this dataset are conditioned on the neutral **theme** label.
+- **Frames:** 384×216 RGB JPEG.
+- **Actions:** one per frame, 5-dim binary multi-hot `[accelerate, brake, left, right, drift]`.
+- **Layout:** each `data/<theme>/<clip>.tar` unpacks to `<clip>/frames/frame_*.jpg` and `<clip>/actions.jsonl` (one JSON line per frame: `{frame_file, actions, theme, game}`).
+- **`metadata.jsonl`:** clip index, one line per clip: `{clip, theme, game, n_frames, path}`.