Datasets:
Leaderboard submission: Goddragon/Arbiter-4B
Model: https://huggingface.co/Goddragon/Arbiter-4B
Mode: general, fine-tuned on train, temperature-calibrated
Training: Initialized from TokenRhythm/NeoHorse-Jev-4B existing weights, then LoRA-fine-tuned on our collected decision datasets, followed by global temperature calibration. The project collection includes typed-decisions train, direct classification datasets, TRAM, HelpSteer2 and constructed examples; exact run settings, when present, are retained in model_manifest.json. The project recipe separates validation from training and reserves typed-decisions test for final evaluation. Calibration uses held-out validation soft labels; it does not retrain the pointer head. This is not a zero-shot submission.
Test: 400 cases / 2,000 decisions; five questions per request.
| Metric | Value |
|---|---|
| accuracy | 0.817000 |
| kl | 0.092212 |
| brier | 0.047905 |
| ece_15 | 0.183987 |
KL(gold || prediction), natural log; Brier sums squared errors against soft gold over classes; ECE uses 15 equal-width top-label bins.
Hardware: A800
p50 latency: 132.923 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).
Report and predictions: https://huggingface.co/Goddragon/Arbiter-4B/tree/main/evaluation
These results are self-reported. Please consider listing this model in the appropriate table.