Datasets:
Qwen3.5-4B LoRA fine-tune: results submitted, model missing from the Hub leaderboard
Hi, I fine-tuned unsloth/Qwen3.5-4B on the train split (LoRA r=16 plus a Clef decision
head, trained with Unsloth) and added results to my model repo:
https://huggingface.co/wozniakos10/Qwen3.5-4B-typed-decisions
Mode: fine-tuned on train, so not zero-shot.
Test split, config all, 400 cases / 2,000 decisions, one request per case with all five questions.
| Accuracy | KL from gold | Brier | ECE |
|---|---|---|---|
| 0.7725 | 0.0992 | 0.0543 | 0.1387 |
KL is KL(gold || prediction) and Brier is summed over options. My scorer reproduces the
Uniform reference row (KL 0.444, Brier 0.238). ECE is top-label with 10 equal-width bins.
Per-decision predictions are in eval_predictions.jsonl in the repo.
The Hub parses .eval_results/typed-decisions.yaml (all four tasks show up in the model API),
but the model does not appear on the benchmark leaderboard or in the model page widget.
Do you know what else is required? I would also be glad to be added to the fine-tuned table.