Add model card

README.md

@@ -0,0 +1,36 @@
+---
+library_name: pytorch
+tags:
+  - world-model
+  - tokenizer
+  - autoencoder
+  - dinov3
+  - counter-strike
+  - opencs2
+---
+
+# CS2 RAE — DINOv3 representation autoencoder tokenizer (d32, 128px)
+
+Frozen latent tokenizer for the [OpenCS2 world model](https://github.com/julien-blanchon/world-model).
+It turns each `128×128` game frame into a tiny `32×8×8` latent (a `8×8 = 64`-token grid),
+and the world model lives entirely in that latent space.
+
+- **Architecture:** DINOv3-based representation autoencoder (RAE) with a learned bridge and a
+  subpixel decoder (1 conv + PixelShuffle), trained natively at 128px.
+- **Latent:** `32×8×8` (`latent_channels=32`, `latent_size=8`) — 192× spatial compression.
+- **Training:** full 3-stage recipe (stage1 → stage2 → stage3) with LPIPS + late GAN; seam-free
+  subpixel decoder. This is the **stage-3 final** checkpoint and the default tokenizer for the
+  world model (`runs/rae/full_d32_128/stage3.pt`).
+
+## Files
+
+| File | Description |
+|---|---|
+| `stage3.pt` | Stage-3 final RAE weights (encoder + bridge + decoder). |
+
+## Usage
+
+Load with the [`world-model`](https://github.com/julien-blanchon/world-model) codebase
+(`RAEConfig` with `latent_channels=32`). The world model consumes this tokenizer frozen.
+
+> Internal research checkpoint. Provided as-is, no warranty.