mR3-Qwen3-8B-en-prompt-en-thinking Overview

mR3-Qwen3-8B-en-prompt-en-thinking is an 8 billion parameter reward model, part of the Multilingual Rubric-Agnostic Reward Reasoning Models (mR3) family. It is fine-tuned from the Qwen3-8B base model and specializes in evaluating assistant responses against detailed rubrics, providing both a score and corresponding reasoning. The model has a context length of 32768 tokens.

Key Capabilities

• Multilingual Evaluation: Trained on a diverse dataset covering 72 languages, enabling robust evaluation across a wide linguistic spectrum.
• Rubric-Agnostic Reasoning: Designed to provide reasoning and scores based on various evaluation rubrics, including factors like safety, helpfulness, relevance, conciseness, politeness, and coverage.
• Task Versatility: Applicable to a range of tasks such as classification, preference optimization, and question answering.
• Detailed Feedback: Generates an explanation comparing responses and a clear verdict (e.g., 'Assistant A' or 'Assistant B').

Good For

• Automated Content Moderation: Evaluating the safety and appropriateness of generated text.
• Response Quality Control: Assessing the helpfulness, relevance, and overall quality of AI assistant outputs.
• Preference Optimization: Providing structured feedback for training and refining language models.
• Multilingual Applications: Evaluating responses in a broad array of languages, leveraging its 72-language training data.

For more technical details, refer to the mR3 paper.