rubricreward/mR3-Qwen3-14B-en-prompt-en-thinking
rubricreward/mR3-Qwen3-14B-en-prompt-en-thinking is a 14 billion parameter reward model from the mR3 family, fine-tuned from Qwen3-14B. It is designed for Multilingual Rubric-Agnostic Reward Reasoning, performing Supervised Fine-Tuning (SFT) on a curated dataset covering 72 languages. This model excels at tasks like classification, preference optimization, and question answering by providing scores and reasoning based on evaluation rubrics.
Loading preview...
Model Overview
mR3-Qwen3-14B-en-prompt-en-thinking is a 14 billion parameter reward model, part of the mR3 (Multilingual Rubric-Agnostic Reward Reasoning) series. Developed by rubricreward, this model is fine-tuned from the Qwen3-14B architecture using Supervised Fine-Tuning (SFT) on a unique dataset. The dataset comprises examples from 72 languages, covering tasks such as classification, preference optimization, and question answering. Each entry includes an instruction, task description, input, response(s), evaluation rubrics, a score, and corresponding reasoning in both English and non-English.
Key Capabilities
- Multilingual Reward Reasoning: Trained on a diverse dataset spanning 72 languages, enabling robust evaluation across various linguistic contexts.
- Rubric-Agnostic Evaluation: Designed to provide scores and reasoning based on flexible evaluation rubrics, making it adaptable to different assessment criteria.
- Task Versatility: Proficient in evaluating responses for classification, preference optimization, and question answering tasks.
- Detailed Reasoning: Generates an explanation and a verdict (e.g., 'Assistant A' or 'Assistant B') based on a systematic comparison of factors like safety, helpfulness, relevance, conciseness, politeness, and coverage.
Usage and Integration
This model can be effectively used with popular libraries like Hugging Face's transformers and vLLM for faster inference. It supports a thinking mode, which can be enabled during generation to provide more nuanced reasoning. The model is licensed under Apache 2.0, and further details can be found in the associated paper.