metadata
license: apache-2.0
Model Card for Model ID
This model improves the instruction-following capabilities of Qwen-2.5-3B-Instruct via preference tuning on the WildChecklists dataset. Full details are provided in Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning.
Evaluation
We report performance on instruction-following and general-chat benchmarks, using GPT-5-mini as the judge. Additional evaluation details and settings are provided in the paper. Alpacaeval/Arena-Hard:
| Model | Alpacaeval (Vanilla) | Alpacaeval (Length-Controlled) | Arena-Hard (Vanilla) | Arena-Hard (Style-Controlled) |
|---|---|---|---|---|
| Qwen-2.5-3B-Instruct | 24.1 | 12.79 | 20.6 | 19.6 |
| + PROSPER | 35.8 | 20.94 | 26.0 | 24.6 |