Datasets:
ZhiPan2-4B — specialist (fitted on train), closed weights, predictions public
Submitting ZhiPan2-4B for the fitted/specialist table.
Results (official scorer formulas, test split, 400 cases / 2,000 decisions)
| accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ |
|---|---|---|---|
| 0.7710 | 0.1105 | 0.0587 | 0.1625 |
Per-primitive top-1: choice 67.33 (n=600) · noul 86.33 (n=600) · score 77.50 (n=800)
Model. Qwen3.5-4B backbone + typed decision head with per-primitive temperature
calibration (temps 1.0/1.2/1.0), trained on the train split only (per-type mixed
distillation: score pure-gold + choice/noul soft targets, τ=0.6 gating). 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/zhipan2-4b-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 publishing the predictions. I rescored them with bench/score_with_jobs.py::score_one against the test split. The file covers all 2,000 decisions (400 cases), with no missing or extra keys, and the numbers match yours: accuracy 0.7710, KL 0.1105, Brier 0.0587, ECE 0.1625. Per type: choice 67.33, noul 86.33, score 77.50.
ZhiPan2-4B is now in the fitted table at rank 25, linked to this thread and the predictions dataset: https://huggingface.co/datasets/LocalLLaMA/typed-decisions/commit/13013327698b4710d84f7c9580ca48d5568a39c7