Training (verified)
- Data:
joint_v0.43 — 91,533 train / 5,085 valid rows (bosnian-core, knowledge, RAG, function-calling, agentic-terminal, coding, safety, ...)
- 1 epoch, 2,042 steps, max_len 2048, LoRA r=64 alpha=128, bf16
- Runtime: 15h23m; train_loss 0.6174
- Validation (200 rows): eval_loss 1.1529, mean_token_accuracy 76.18%
Honest limitations (measured 2026-11-05)
- Tool-calling eval (63 GAMA runtime tasks): tool_acc 0.0%, args_acc 3.7%, parse_rate 49.2%
- Cause: training data contains only 276 tool-call rows (0.3%), and none of the GAMA runtime tools (
hc, sql, brain_search, gama_run, doc, web_search); model learned generic service_status/package_install names and over-refuses benign requests.
- Not suitable for production tool calling. Intended for BCMS text/RAG/coding; tool schema retraining planned (v2).
Files
adapter_config.json, adapter_model.safetensors (LoRA), tokenizer, chat_template.
Base model: https://huggingface.co/Qwen/Qwen3.5-9B