qwen3-1.7b-digestive-coach-n1102

Digestive Wellness Coach — Qwen3-1.7B, N=1102 (headline checkpoint, T10d round 2)

Unsloth QLoRA fine-tune of Qwen/Qwen3-1.7B on 1102 rows of kelsbeans/digestive-coach-dataset (nested subset of one shuffle; fixed family-disjoint valid set of 81). Merged fp16 — loads with plain transformers, no peft. Adapter: kelsbeans/qwen3-1.7b-digestive-coach-n1102-adapter.

Pinned for verification (R48/RUBRIC-09)

• Weights revision (pre-card commit): ffc4be4619c247b4ca5a2cb7d84ea5cd00faed8c
• Eval-code git hash: 75863851837a7cc853b7052704e8ef7b6f7f8690
• Best eval loss 0.7831 at step 276; early stopping on val loss

Eval (held-out 40-scenario family-disjoint set, greedy, enable_thinking=false, max_new_tokens=1600, seed 7)

Judge means (tuned-only run, anthropic/claude-sonnet-5, rubric judge-v1, 1-5)

integrated_voice 4.58 · honest_uncertainty 4.30 · no_manufactured_agreement 4.36 (33/40 usable). Robustness 1/15 (first nonzero at any N); buried-red-flag recall 2/5 vs 0/5 at n781 — see docs/t10d-diagnosis.md.

Reproduce: python eval.py --model kelsbeans/qwen3-1.7b-digestive-coach-n1102 --base Qwen/Qwen3-1.7B --eval-set eval_scenarios.jsonl --max-new-tokens 1600 (harness in the project repo at the pinned eval-code hash; raw per-example transcripts ship as JSONL in runs/).

Best prompted frontier cell on the same eval: 15% spec adherence — this checkpoint (25%) and n781 (20%) are the two points above that ceiling.

Not medical advice. Research artifact for a data→behavior fine-tuning project.