rotating-equip-sft

Rotating Equipment Evidence-Gated QLoRA

Qwen3-1.7B QLoRA specialized for one falsifiable rule: diagnose only when two independent evidence channels agree; otherwise WATCH without naming a mode. Responses must identify the asset, cite independent evidence, and give exactly one non-compound action.

Corrected N=2580 result

• Full-Spec adherence / robustness: 0.933 / 0.933 (28/30)
• Exact GATE line: 30/30
• Thinking dumps: 0/30
• Remaining misses: e11, e14 watch over-calls
• Adapter revision: 784a90897112cd94effb2da39f2fc151f55468ae
• Training evidence: repository commit 7bab0c5, results/training/n2580/

This is below the declared 0.95 bar and is not a threshold pass.

Exact inference

The adapter was trained against the listed bitsandbytes 4-bit base. Exact adapter evaluation requires CUDA. Applying the LoRA weights to upstream full-precision Qwen on CPU/MPS changes outputs and is non-comparable.

python eval.py \
  --model lecporr/rotating-equip-sft \
  --revision 784a90897112cd94effb2da39f2fc151f55468ae \
  --eval-set data/eval.jsonl \
  --device cuda

A merged 16-bit checkpoint is required for portable CPU/MPS one-command inference; see notebooks/COLAB_MERGE.md.

Limitations

Synthetic snapshots only. This model is not a physical diagnostic authority or safety controller. The 30-row public eval is small; staff-held-out evaluation remains required.