Upload README.md with huggingface_hub

README.md

@@ -1 +1,31 @@
-# CWM-Extended adapters (uploading...)
+---
+license: other
+license_name: fair-noncommercial-research
+base_model: facebook/cwm
+tags: [lora, peft, code-world-model, world-model]
+---
+
+# CWM-Extended LoRA adapters
+
+Trained LoRA adapters (r=16, attn+MLP) for the **CWM interactive/visual world-model** project
+(code + full REPORT: https://github.com/namak-kun/cwm-extended). Base model: `facebook/cwm`.
+
+Each subfolder is a loadable PEFT adapter. Headline ones:
+
+| adapter | what it teaches | result |
+|---|---|---|
+| `cwm_gametick_stepover` | one-shot game-tick transition | 0.017 -> 0.692 |
+| `cwm_fdm_idm_r1` / `_r2` | FDM<->IDM self-labeling flywheel | 0.525 -> 0.683 (~= oracle), stable 2 rounds |
+| `cwm_cascade` | UI DOM-cascade step-over | in-dist win (uidom 0.80->1.0); NB negative transfer to real JS |
+| `cwm_fdm_oracle_r1/_r2`, `cwm_fdm_hardoracle` | oracle-label controls | flywheel validation |
+| `cwm_heldapp` | UI minus togglelist (abstraction test) | cross-app 0.44 -> 0.56 |
+| `cwm_oop_expanded`, `cwm_mixed_expanded` | phi-expansion + forgetting fix | 0.02 -> 0.93 |
+| `cwm_arith_*`, `cwm_dagger_*` | arithmetic / DAgger studies | mostly neutral (capability hole) |
+
+Load (PEFT):
+```python
+from peft import PeftModel
+from transformers import AutoModelForCausalLM
+base = AutoModelForCausalLM.from_pretrained("facebook/cwm")
+model = PeftModel.from_pretrained(base, "nmk-kun/cwm-extended-adapters", subfolder="cwm_gametick_stepover")
+```