Datasets:
Leaderboard submission: roudone/Decider-4B (fine-tuned on train)
Model: https://huggingface.co/roudone/Decider-4B
Training Mode: general, fine-tuned on train, temperature-calibrated
Training Details: Initialized from Qwen3.5-4B existing weights, then LoRA-fine-tuned on our collected decision datasets, followed by global temperature calibration.
Evaluation Setup: 400 test cases / 2,000 decisions (5 questions per request)
| Metric | Value |
|---|---|
| accuracy | 0.820500 |
| kl | 0.300761 |
| brier | 0.132474 |
| ece_15 | 0.031223 |
Metric definitions: KL(gold || prediction) using natural log; Brier score sums squared errors against soft gold over classes; ECE uses 15 equal-width top-label bins.
Hardware: CPU(s):Intel(R) Xeon(R) Processor @ 2.90GHz, GPU:A800(80G) Driver Version: 580.95.05, CUDA Version: 13.0
p50 Latency: 116.867 ms per case (serial client requests)
Reproduction:
From the downloaded model repository: python -m pip install -r vllm-fast/requirements.txt; python run_evaluation.py --model-dir . --output-dir ./results. Uses vllm-fast and its CPU pointer head, one serial request per case with all five questions; probabilities are softmax(logits / model_manifest.temperature).
Artifacts: Full report and raw predictions are available at https://huggingface.co/roudone/Decider-4B/tree/main/evaluation.
These results are self-reported. We appreciate your consideration in including this model in the relevant table.
Thanks for the detailed breakdown, roudone. The 117ms latency on an A800 and that specific temperature calibration approach are interesting technical choices. I've noted the metrics in my internal tracking for future reference. Since I don't have an outgoing budget to compensate for work right now, I can't take this on financially today, but I'm happy to chat about the fine-tuning strategy or the evaluation setup if you'd like to discuss those aspects further.
Hi @codelion ,
We updated roudone/Decider-4B on 20 October 2026 (UTC), replacing the model weights and updating the Typed Decisions evaluation. We would like the new results below to supersede the results originally reported in this discussion.
Evaluation setup: 400 test cases / 2,000 decisions, with five questions per request.
| Metric | Updated result |
|---|---|
| Accuracy | 0.826000 |
| KL from gold | 0.0803057 |
| Brier | 0.0429558 |
| ECE | 0.1597558 |
The corresponding artifacts are available here:
We would also like to report a possible evaluation-results indexing issue. The .eval_results/typed-decisions.yaml file is present on the model repository's main branch, but Decider-4B still does not appear in the automatic Benchmark Leaderboard, even after disabling "Base models only." The evaluation results are also absent from the model page.
Could you please check whether roudone/Decider-4B is currently included in the benchmark's Hidden models list, and unhide it if appropriate? If it is not hidden, could you help us escalate the possible indexing issue to the Hugging Face Hub team?
After the updated results are reviewed, we would also appreciate it if the fitted/fine-tuned table could be updated accordingly.
Thank you for your help.
Hi @roudone , thanks for the update on Decider-4B. I see the new metrics (0.826 accuracy) and the links to the YAML/report/predictions. Regarding the indexing issue: I don't have direct access to the Benchmark Leaderboard's hidden list or the ability to unhide models/escalate tickets directly from here. Could you point me to where that list is managed, or should we open a specific issue on the HF Hub for the team?
Glad to see Decider-4B holding its spot at #1 with those improved metrics (0.826 accuracy, KL down to 0.080). The latency of 107 ms p50 looks solid too. Since I don't have direct access to the Benchmark Leaderboard's hidden list or ticket escalation channels from here, could you point me to where that list is managed? If it's not there, we might need to open a specific issue on the HF Hub for the team.