Datasets:
ZhiPan3-4B — specialist (fitted on train), closed weights, predictions public
Submitting ZhiPan3-4B for the fitted/specialist table.
Results (official scorer formulas, test split, 400 cases / 2,000 decisions)
| accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|
| 0.8005 | 0.0967 | 0.0503 | 0.1598 |
Per-primitive top-1: choice 74.33 (n=600) · noul 87.67 (n=600) · score 78.62 (n=800)
Model. Qwen3.5-4B backbone + typed decision head with per-workflow per-primitive
calibration (log-space gamma/bias with a direction-matching prior), trained on thetrain split only. 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/zhipan3-4b-td-predictions — re-scorable withbench/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.
Naming correction: this submission should be ZhiPan2.1-4B (not ZhiPan3-4B).
ZhiPan2.1 = same architecture/training lineage as ZhiPan2 (r26) + per-workflow per-primitive
calibration — an iteration, not an architecture change. The README has been updated accordingly.
Self-reported: accuracy 0.8005 / KL 0.0967 / Brier 0.0503 / ECE 0.1598.
Thanks for the predictions and the naming note. ZhiPan2.1-4B is now in the fitted table at #10 of 50. 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 0.8005, KL 0.0967, Brier 0.0503, ECE 0.1598; choice 74.33, noul 87.67, score 78.62). It ties pramana-2b on accuracy and sits just below it, since pramana-2b has the lower KL and Brier. The ZhiPan2-4B row stays as a separate entry, now at #26.