Datasets:
TypeCraft ens3 (specialist, train only): accuracy 0.821 · KL 0.0708 · Brier 0.0388 · ECE 0.183
TypeCraft ens3 — specialist, fitted on train only: accuracy 0.821 · KL 0.0708 · Brier 0.0388 · ECE 0.183
Authors: Yifan Yang, Nicole Chu
Mode: specialist (fitted on this benchmark's train split; not comparable with zero-shot rows). No external decision data, no teacher/LLM labels — train gold distributions only.
Scores on test (400 cases, 2000 decisions, full distributions), computed with the dataset's own bench/score_with_jobs.py scorer:
| flavour | accuracy | KL from gold | Brier | ECE |
|---|---|---|---|---|
| submitted (calibrated) | 0.821 | 0.0708 | 0.0388 | 0.183 |
| same ensemble, uncalibrated | 0.8155 | 0.0717 | 0.0393 | 0.1685 |
What it is. A probability average of three per-question "decision-head" scorers — Qwen3-1.7B (full fine-tune), Qwen3-4B (LoRA r32, one round of a member-swap self-improvement loop) and Qwen2.5-3B-Instruct (LoRA r16). Each reads the state, the question and every option's criteria text, scores the options with a bilinear head over the option tokens' hidden states, and is trained with cross-entropy against the gold distribution. The average is then calibrated with one temperature and one bias per (workflow, question, option), fitted on 200 train cases held out from all training.
How it was selected. Members, ensemble and calibration were all chosen on that 200-case held-out slice of train; test was used once, for the final score. Train and test were checked disjoint (ids and states), and the shipped checkpoints reproduce these predictions bit-exactly.
Caveats, stated up front.
- The calibration's +0.55pt accuracy over the uncalibrated ensemble is not supported by a cross-fitted re-analysis on the held-out slice, so read it as noise-level; the calibration does lower KL and Brier.
- Bit-exact on the A100 that produced it; a re-run on different accelerator hardware gave 0.8185 (99.4% of argmax decisions identical).
Predictions file sha256: bba6b6f5a7dda3f638b4f54ff9e0c51a80be74a147635a7d4252d5a35a197e43 (2000 rows: id, workflow, question, type, probabilities); weights and predictions are not public yet and can be shared for rescoring on request.
Predictions on git https://github.com/yfyang86/TypeCraft_ens3/blob/main/predictions.jsonl
Thanks for sharing the predictions. I rescored predictions.jsonl with score_one from bench/score_with_jobs.py against the test split: all 2,000 decisions are covered and the numbers match yours exactly (accuracy 0.821, KL 0.0708, Brier 0.0388, ECE 0.183; choice 0.797, noul 0.868, score 0.804). TypeCraft ens3 is now in the fitted table at #4 of 53, just below Decider-4B (0.826), and it has the lowest KL and Brier of any row in that table: https://huggingface.co/datasets/LocalLLaMA/typed-decisions