Upload README.md

README.md

@@ -1,6 +1,6 @@
 ---
-title: RFAB Historic Chat Space - GPT 1900
-emoji: 👴
+title: RFAB Historic Chat Space
+emoji: 🕰️
 colorFrom: blue
 colorTo: purple
 sdk: gradio
@@ -8,7 +8,7 @@ sdk_version: 5.49.1
 app_file: app.py
 pinned: false
 models:
-  - mhla/gpt1900-instruct-v3-sft
+  - Pclanglais/MonadGPT
 ---
 
 # RFAB Historic Chat Gradio Space Template
@@ -60,7 +60,10 @@ GPT-1900:
 
 ```bash
 RFAB_HISTORIC_MODEL_ID=mhla/gpt1900-instruct-v3-sft
-RFAB_HISTORIC_ADAPTER=transformers
+RFAB_HISTORIC_ADAPTER=nanochat
+RFAB_NANOCHAT_MODEL_FILE=model_000075.pt
+RFAB_NANOCHAT_META_FILE=meta_000075.json
+RFAB_GPU_DURATION_SECONDS=240
 ```
 
 Mr. Chatterbox:
@@ -70,9 +73,19 @@ RFAB_HISTORIC_MODEL_ID=tventurella/mr_chatterbox_model
 RFAB_HISTORIC_ADAPTER=nanochat
 RFAB_NANOCHAT_MODEL_TAG=d18
 RFAB_NANOCHAT_SOURCE=sft
+RFAB_NANOCHAT_MODEL_FILE=model_000117.pt
+RFAB_NANOCHAT_META_FILE=meta_000117.json
+RFAB_GPU_DURATION_SECONDS=240
 ```
 
-The `nanochat` adapter is scaffolded for the Mr. Chatterbox code path. The Space must include the same nanochat package/checkpoint layout as `tventurella/mr_chatterbox` before `DRY_RUN=false`.
+The `nanochat` adapter lazily downloads the checkpoint, loads `tokenizer/`, `meta_*.json`, and `model_*.pt`, then generates with nanochat special tokens. Keep `RFAB_GPU_DURATION_SECONDS` high on first real smoke because the first request downloads and loads the checkpoint.
+
+The template clones nanochat at runtime instead of installing it with pip because the upstream repo is not currently pip-wheel-installable from GitHub. Optional runtime override:
+
+```bash
+RFAB_NANOCHAT_REPO=https://github.com/karpathy/nanochat.git
+RFAB_NANOCHAT_REF=1076f97
+```
 
 ## Hardware