loop-qwen35-4b — Qwen3.5-4B closed-loop insulin controller (Gemini-distilled)

A Qwen3.5-4B model LoRA-distilled from a gemini-3-flash-preview insulin-control policy for closed-loop type-1-diabetes control in the simglucose simulator. It emits the next basal/bolus action chunk (with a short clinical rationale) from a rolling CGM/insulin/carb history + a deterministically-computed insulin-on-board.

Result (held-out 9 patients, 48 h, seed 42)

meals	model	TIR (70–180)	TBR (<70)	survived
announced	v15 (Qwen3-4B)	68.8	13.2	7/9
announced	this (Qwen3.5-4B)	80.8	2.1	9/9
unannounced	v15	68.9	10.3	8/9
unannounced	this	73.8	2.2	9/9
unannounced	Gemini teacher	73.9	3.0	9/9

• 9/9 survival in both conditions — matches the Gemini teacher; unannounced TIR ≈ teacher parity. The base-model upgrade from Qwen3-4B (v15) collapsed the pediatric over-dosing (TBR 13/10 → ~2 %) and saved the hardest patient (child#008).
• Eval is bf16 on A100 (single seed); Q4_K_M deployment verified surviving on the decisive hard case.

Files

• adapter_model.safetensors — LoRA adapter (rank 32, all-linear targets; the hybrid linear-attention layers need all-linear). Merge onto Qwen/Qwen3.5-4B.
• *-f16.gguf — merged f16 GGUF, already patched for llama.cpp/Ollama (Qwen3.5's phantom MTP block removed: block_count −1, nextn_predict_layers=0). Quantize + serve: ollama create loop-qwen35 --quantize q4_K_M -f Modelfile (needs Ollama ≥ 0.32). Modelfile: FROM *-f16.gguf / TEMPLATE {{ .Prompt }} / PARAMETER num_ctx 4096 / PARAMETER temperature 0.

Training

transformers 5.x + torch 2.6 (matched torchvision/torchaudio), LoRA r32 target_modules=all-linear, 2 epochs, eff. batch 32, lr 2e-4, max-seq 4096. Data: jxx123/loop-qwen-v8-sft (distill_sft_v16.jsonl). Reason-before-act (concise clinical rationale) output format. Trained on 21 non-held-out patients only.

⚠️ Research / simulation only — not a medical device.