Datasets:
Submission: rony/ministral3-3b-reasoning-typed-decisions (Specialist) (Ministral 3 3B Reasoning, fine-tuned on train)
Model: https://huggingface.co/rony/ministral3-3b-reasoning-typed-decisions
Mode: (Specialist) fine-tuned on train (all four workflows), for the "Fitted or fine-tuned on train" table.
Scored on LocalLLaMA/typed-decisions all/test (400 cases, 2000 decisions), never seen in training.
Machine: DGX Spark
| Model | Kind | Accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ | Latency |
|---|---|---|---|---|---|---|
| [ministral3-3b-reasoning-typed-decisions](httpuv run jupyter lab ministral3_3b_reasoning_decisions.ipynbs://huggingface.co/rony/ministral3-3b-reasoning-typed-decisions) (Specialist; Ministral 3 3B Reasoning, full fine-tune) | fine-tuned on train |
0.808 | 0.0773 | 0.0424 | 0.1822 | 373 ms / case (NVIDIA GB10) |
ECE is computed per question and averaged over the 20 questions (closest to the README's reference rows); pooled over all decisions it is 0.1742.
By question type
| accuracy | kl | brier | ece | |
|---|---|---|---|---|
| choice | 0.803 | 0.103 | 0.053 | 0.204 |
| noul | 0.853 | 0.039 | 0.030 | 0.102 |
| score | 0.779 | 0.086 | 0.044 | 0.207 |
| OVERALL | 0.808 | 0.077 | 0.042 | 0.174 |
By workflow
| accuracy | kl | brier | ece | |
|---|---|---|---|---|
| agent_trace_observability | 0.748 | 0.081 | 0.049 | 0.186 |
| customer_service | 0.836 | 0.086 | 0.043 | 0.188 |
| invoice_processing | 0.866 | 0.078 | 0.042 | 0.122 |
| security_incidents | 0.784 | 0.065 | 0.036 | 0.206 |
| OVERALL | 0.808 | 0.077 | 0.042 | 0.174 |
Baselines on the same split: Prior accuracy 0.478 / KL 0.327; zero-shot Ministral 3 3B Reasoning accuracy 0.340 / KL 1.659.
Setup. Full fine-tune of the text decoder of mistralai/Ministral-3-3B-Reasoning-2512 (vision tower dropped, tied embedding / LM head
frozen, reasoning disabled with an empty [THINK][/THINK] block; AdamW, lr 3.13e-06, weight decay 0.029, warmup 0.07,
effective batch 16, 10 epochs; hyperparameters from Optuna search (hpo/ministral3-3b-reasoning/best_hparams.json); epoch picked on
held-out-train KL).
Each question is its own forward pass (5 per case, not all five in one request); the answer distribution is the softmax
of the next-token logits over the option letters. ECE is 10-bin expected calibration error of max probability vs.
correctness. Full test distributions are in test_predictions.jsonl in the model repo, so the numbers can be re-run.
Added to the fitted-on-train table at 0.808 / KL 0.077, marked as self-reported. Thanks for the per-workflow breakdown and the prediction dumps.