reward-model
OpenJudge: A Unified Framework for Holistic Evaluation and Quality Rewards
Pairwise LLM judges (A/B/tie): budget-aware multi-turn packing, position-bias correction, pseudo-label distillation. Generalized from the 4th-place (gold) solution to Kaggle LMSYS Chatbot Arena.
Light, Efficient, Omni-modal & Reward-model Driven Reinforcement Fine-Tuning Framework
