Add LoRA training scripts and fix bake-off token budget

- training/scripts/train_lora.py: Unsloth QLoRA trainer for qwen3:8b
- training/scripts/train_lora.sh: Launch script for steel141 RTX 3090 Ti
- eval/bakeoff.py: Fixed token budget (400->1500) that caused qwen3
  models to exhaust tokens on thinking, added --no-think flag
- agent/serve.py: Default model changed to gemma3n:e4b

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

eval/bakeoff.py

@@ -31,7 +31,8 @@ RESULTS_DIR = ROOT / "eval" / "results"
 
 
 def ollama_chat(model: str, messages: list, ollama_url: str,
-                temperature: float = 0.2, max_tokens: int = 400) -> dict:
+                temperature: float = 0.2, max_tokens: int = 1500,
+                no_think: bool = False) -> dict:
     """Call Ollama and return response + timing."""
     payload = {
         "model": model,
@@ -43,6 +44,12 @@ def ollama_chat(model: str, messages: list, ollama_url: str,
             "num_predict": max_tokens,
         },
     }
+    if no_think:
+        # Prepend /no_think to the last user message to disable thinking tokens
+        for msg in reversed(payload["messages"]):
+            if msg["role"] == "user":
+                msg["content"] = "/no_think\n" + msg["content"]
+                break
     start = time.time()
     r = requests.post(f"{ollama_url}/api/chat", json=payload, timeout=180)
     r.raise_for_status()
@@ -157,7 +164,7 @@ def score_result(example: dict, actual_cmds: list, parsed: dict) -> dict:
     }
 
 
-def run_bakeoff(models: list, ollama_url: str):
+def run_bakeoff(models: list, ollama_url: str, no_think: bool = False):
     """Run all models against the dataset and compare."""
     # Load dataset
     with open(DATASET) as f:
@@ -166,6 +173,8 @@ def run_bakeoff(models: list, ollama_url: str):
     print(f"Bake-off: {len(examples)} examples × {len(models)} models")
     print(f"Ollama: {ollama_url}")
     print(f"Models: {', '.join(models)}")
+    if no_think:
+        print("Mode: /no_think (thinking tokens disabled)")
     print("=" * 70)
 
     all_results = {}
@@ -208,7 +217,7 @@ def run_bakeoff(models: list, ollama_url: str):
 
             # Call LLM
             try:
-                resp = ollama_chat(model, messages, ollama_url)
+                resp = ollama_chat(model, messages, ollama_url, no_think=no_think)
             except Exception as e:
                 print(f"  [{i+1}/{len(examples)}] ERROR: {e}")
                 results.append({"id": eid, "error": str(e)})
@@ -311,9 +320,11 @@ def main():
     parser.add_argument("--ollama-url", default="http://192.168.0.179:11434")
     parser.add_argument("--models", nargs="+",
                         default=["qwen3-coder:30b", "gemma3n:e4b"])
+    parser.add_argument("--no-think", action="store_true",
+                        help="Prepend /no_think to disable thinking tokens (helps Qwen models)")
     args = parser.parse_args()
 
-    run_bakeoff(args.models, args.ollama_url)
+    run_bakeoff(args.models, args.ollama_url, no_think=args.no_think)
 
 
 if __name__ == "__main__":