Serve fine-tuned ttw-trader-0.5b on the Space (Well-Tuned + tinier Tiny Titan)

- serve.py: bake TTW_MODEL/TTW_MODEL_REVISION into the image env so the chosen
model actually reaches the container (a local deploy-time env var does not
propagate into Modal on its own). Deployed with TTW_MODEL=AdmiralTaco/ttw-trader-0.5b.
- README: document the fine-tuned 0.5B (distilled from cleaned 3B traces;
0% self-buy, 100% valid offers, one-sixth the size) and link the model.
- .gitignore: ignore generated scripts/finetune/sft.jsonl.

Live endpoint verified serving the tuned model: self_buy 0%, valid_offer 100%.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitignore

@@ -5,6 +5,9 @@
 # Generated trace artifact (published to a separate HF dataset, not the app repo)
 traces.jsonl
 
+# Generated fine-tuning set (built from traces; not part of the app repo)
+scripts/finetune/sft.jsonl
+
 __pycache__/
 *.pyc
 .pytest_cache/

README.md

@@ -19,13 +19,17 @@ widening wealth gap emerge on their own.
 
 ## Read more
 - 📓 **Field notes** (what I built and what I learned): [Hugging Face article](https://huggingface.co/blog/build-small/thousand-token-wood-sim) | [Medium](https://medium.com/@LesterLeong/thousand-token-wood-emergent-market-drama-from-3-billion-parameter-agents-22545d5982bf)
+- 🧠 **Fine-tuned model** powering the agents: [AdmiralTaco/ttw-trader-0.5b](https://huggingface.co/AdmiralTaco/ttw-trader-0.5b)
 - 📡 **Open agent traces** on the Hub: [dataset](https://huggingface.co/datasets/build-small-hackathon/thousand-token-wood-traces)
 
 ## Why "small" is load-bearing
 A living economy needs *many* agents thinking *many* times. Frontier models are
-too slow and costly for that. A 3B model makes a real-time multi-agent simulation
-possible, and that is the whole point. The agents run on **Qwen2.5-3B-Instruct**,
-served with vLLM on **Modal**; this Gradio app is just the window onto the wood.
+too slow and costly for that, so the whole thing runs on a **fine-tuned 0.5B
+model** ([ttw-trader-0.5b](https://huggingface.co/AdmiralTaco/ttw-trader-0.5b)).
+It was distilled from cleaned traces of a 3B teacher (self-buy mistakes stripped),
+so a model one-sixth the size trades *more* cleanly than its teacher: zero self-buys,
+100% valid offers. Served with vLLM on **Modal**; this Gradio app is the window
+onto the wood.
 
 ## What you can do
 - **Step / Auto-run** the simulation and watch the market move.

serve.py

@@ -24,7 +24,17 @@ image = (
         "vllm==0.6.6",
         "huggingface_hub[hf_transfer]==0.26.2",
     )
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1", "VLLM_DO_NOT_TRACK": "1"})
+    # Bake the chosen model into the image env so the CONTAINER sees it. A local
+    # `TTW_MODEL=... modal deploy serve.py` env var does NOT propagate into the
+    # Modal container on its own (it re-imports fresh), so capture it here.
+    .env(
+        {
+            "HF_HUB_ENABLE_HF_TRANSFER": "1",
+            "VLLM_DO_NOT_TRACK": "1",
+            "TTW_MODEL": MODEL,
+            "TTW_MODEL_REVISION": MODEL_REVISION,
+        }
+    )
 )
 
 # Persist the HF cache across cold starts so we download the model once.