Datasets:
| license: apache-2.0 | |
| tags: [type-1-diabetes, closed-loop-insulin, policy-distillation, simglucose] | |
| # loop-distill-data — Gemini→Qwen insulin-control distillation SFT (simulator) | |
| Chat-format SFT (system / user-state / assistant action+reasoning) distilling a | |
| `gemini-3-flash-preview` insulin controller. All states here are from the | |
| **simglucose simulator (virtual patients)** — no real-patient data. IOB is provided | |
| in the prompt; actions = basal/bolus chunks + a concise clinical rationale. | |
| | file | rows | provenance | | |
| |---|---|---| | |
| | `distill_sft_v15.jsonl` | 10,182 | simglucose rollouts + Gemini labels (+ child DAgger) | | |
| | `distill_sft_v16.jsonl` | 10,984 | v15 + Gemini-relabelled announced-child over-dose states | | |
| Held-out patients (adult#007-009, adolescent#008-010, child#008-010) excluded from | |
| training. `distill_sft_v16.jsonl` trained `jxx123/loop-qwen35-4b` (Qwen3.5-4B, | |
| Gemini parity 9/9). Research/simulation only — not a medical device. | |
| Note: real-patient-derived distill data (from de-identified Nightscout traces) is | |
| kept OFF HuggingFace by policy and lives only in the private repo/infra. | |