rubricreward/mR3-Qwen3-4B-en-prompt-en-thinking
rubricreward/mR3-Qwen3-4B-en-prompt-en-thinking is a 4 billion parameter reward model, part of the mR3 (Multilingual Rubric-Agnostic Reward Reasoning Models) family, fine-tuned from Qwen/Qwen3-4B. It is specifically designed for evaluating and reasoning about AI responses across 72 languages, covering tasks like classification, preference optimization, and question answering. This model excels at providing scores and detailed reasoning based on evaluation rubrics, making it suitable for automated content moderation and quality assessment.
Loading preview...
Model Overview
mR3-Qwen3-4B-en-prompt-en-thinking is a 4 billion parameter reward model, developed by rubricreward, and is a member of the mR3 (Multilingual Rubric-Agnostic Reward Reasoning Models) family. This model is fine-tuned from the Qwen3-4B base model and specializes in evaluating AI-generated responses. It leverages a curated mR3 dataset spanning 72 languages, which includes instructions, task descriptions, inputs, responses, evaluation rubrics, scores, and corresponding reasoning.
Key Capabilities
- Multilingual Evaluation: Trained on data from 72 languages, enabling broad application.
- Rubric-Agnostic Reasoning: Provides scores and detailed reasoning for AI responses based on defined rubrics.
- Task Versatility: Supports various tasks including classification, preference optimization, and question answering.
- Detailed Feedback: Generates an explanation and a verdict (Assistant A or Assistant B) based on factors like safety, helpfulness, relevance, conciseness, politeness, and coverage.
Use Cases
- Automated Content Moderation: Assessing the safety and appropriateness of generated text.
- AI Response Quality Assurance: Evaluating the helpfulness and relevance of AI outputs.
- Preference Optimization: Providing structured feedback for training and refining other language models.
- Research in Reward Modeling: A foundational model for exploring multilingual and rubric-based reward mechanisms, as detailed in their paper.
This model is licensed under Apache 2.0 and can be efficiently used with transformers or vLLM for inference.