jxx123/loop-qwen-v8-sft

loop-qwen-v8 SFT dataset (Gemini insulin-control distillation)

8,222 chat-format examples used to SFT loop-qwen-v8 (Qwen3-4B insulin controller distilled from Gemini-3-flash-preview). Each example is a closed-loop dosing decision.

Format (JSONL, one chat per line)

• system: controller spec (IOB-aware, Chain-of-Draft reason-before-act)
• user: patient metadata (age, weight, TDD, CF, IC, basal) + 6h history of CGM / insulin / carbs, as JSON
• assistant: JSON {"reasoning": "...", "basal_chunk": [5], "bolus_chunk": [5]} (reason-first, then the 5-step basal U/hr + bolus U action)

Composition

• Gemini (rules-off + IOB + CoD) demonstrations on the 21 training patients (held-out 9 excluded — leakage guard).
• Multi-seed meals (seeds 42/7/13/21) — the key ingredient that let the 4B student reach teacher parity without dosing oscillation.
• Built via gen_teacher_data.py -> audit_to_sft.py (see the loop-gpt repo).

Companion model

jxx123/loop-qwen-v8 - the 2.5 GB Q4 model trained on this data (held-out-9: TIR 72.4% vs Gemini 73.5%, survival 9/9).

Research artifact

Simulation (simglucose) only. Not a medical device; do not use for real insulin dosing.