Provenance
- Protocol:
qwen3.5-4b-profile-400-sft-v2-r32-cosine-15e-selected-epoch10
- Protocol SHA256:
d3cbada8c6bb2e5e1d676fdf2cf6297c5b93b9a68a4b7f543137dca22a97dabe
- Model fingerprint:
f076ed9c966d6b589b816847bf8abdc0930b50f982ed88f5c230f2d2c5645e24
- SFT selection: complete epoch 10 at optimizer step 2370 from a training run that later completed epoch 12; the cosine scheduler retained its original 15-epoch horizon.
- Evaluation: fixed 4-shot greedy decoding. Classification uses native thinking; mask and generation disable thinking.
The Oracle is evaluated separately and is not counted in the formal six-model matrix.
The repository includes the verified three-direction classification, mask, and generation summary.
See upload_manifest.json, selection/epoch10_selection_manifest.json, training/run_manifest.json, protocol/, and eval/ for machine-readable provenance.