loop-qwen35-2b — Qwen3.5-2B closed-loop insulin controller (Gemini-distilled)

Qwen3.5-2B LoRA-distilled from a gemini-3-flash-preview insulin-control policy for closed-loop type-1-diabetes control in the simglucose simulator. Emits the next basal/bolus action chunk plus a short auditable clinical rationale, from a rolling CGM/insulin/carb history and a deterministically computed insulin-on-board.

Result — held-out 9 patients (48 h, seed 42, bf16)

model	Q4 size	ANN TIR	ANN TBR	ANN surv	UNANN TIR	UNANN TBR	UNANN surv
Qwen3.5-4B	2.7 GB	80.8	2.1	9/9	73.8	2.2	9/9
this (Qwen3.5-2B)	~1.3 GB	81.8	1.8	9/9	73.6	1.6	9/9
Gemini teacher	—	—	—	—	73.9	3.0	9/9

The 2B matches the 4B at half the size — +1.0 pp TIR announced, −0.2 pp unannounced (noise), slightly lower hypo on both, 9/9 survival in both meal conditions with zero deaths. Unannounced TIR 73.6 ≈ teacher parity (73.9). "Unannounced" = 50 % of meals hidden from the controller (patient forgets to announce) — the realistic failure mode. Train loss 0.364 (4B: 0.338).

Files

• adapter_model.safetensors — LoRA (r32, all-linear; Qwen3.5's hybrid linear-attention layers require all-linear). Merge onto Qwen/Qwen3.5-2B.
• qwen35-2b-f16.gguf — merged f16 GGUF, already patched for llama.cpp/Ollama (Qwen3.5's phantom MTP block removed: block_count->24, nextn_predict_layers->0). Deploy: ollama create loop-qwen35-2b --quantize q4_K_M -f Modelfile (Ollama >=0.32), with FROM qwen35-2b-f16.gguf / PARAMETER num_ctx 4096 / PARAMETER temperature 0.

Notes

• Non-thinking: the chat template emits an empty <think></think>; the model answers with JSON directly. Training and eval use identical rendering (no train/serve skew).
• Data: jxx123/loop-distill-data (distill_sft_v16.jsonl, simulator only). Trained on 21 non-held-out patients; the 9 eval patients are excluded.

⚠️ Research / simulation only — not a medical device.