Datasets:
ZhiPan-Z3-Pro — specialist (fitted on train), closed weights, predictions public
Submitting ZhiPan-Z3-Pro for the fitted/specialist table.
Results (official scorer formulas, test split, 400 cases / 2,000 decisions)
| accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|
| 0.8265 | 0.8033 | 0.1515 | 0.0741 |
Per-primitive top-1: choice 77.17 (n=600) · noul 88.83 (n=600) · score 82.12 (n=800)
Model. ZhiPan-Z3-Pro is a per-primitive ensemble (prediction fusion) of 14
fine-tuned runs on a Qwen3.5-4B backbone with typed decision heads — per-type
independent fusion (choice: 6 tracks, noul: 3 tracks, score: 5 tracks) selected by
log-space weight grid search + member add/remove search, all fitted strictly on thetrain split. Weights stay closed; to keep the self-report verifiable we publish
the full test-split predictions (2000 probability distributions) at
https://huggingface.co/datasets/cayouccc/zhipan-z3-pro-td-predictions — re-scorable
with bench/score_with_jobs.py::score_one (noul true/false label mapping included).
Kind: specialist, fitted on train — not comparable with the zero-shot table.
Happy to answer any questions about the recipe.
Thanks for the predictions. ZhiPan-Z3-Pro is now in the fitted table at #1 of 51. I rescored predictions.jsonl with score_one from bench/score_with_jobs.py: all 2,000 decisions are covered, and the numbers match your report exactly (accuracy 0.8265, KL 0.8033, Brier 0.1515, ECE 0.0741; choice 77.17, noul 88.83, score 82.12). ZhiPan2.1-4B and ZhiPan2-4B stay as separate entries, now at #11 and #27.