Model Overview

OpenRubrics/RubricRM-4B-Rubric-v2 is a 4 billion parameter model, fine-tuned from the Qwen3/Qwen3-4B architecture. Its primary function is to generate structured, rubric-style evaluation criteria from user prompts, which are then used to assess the quality of AI-generated responses. This model is a key component in reward modeling and LLM alignment efforts, aiming to provide scalable synthetic rubric generation.

Key Capabilities

• Rubric Extraction: Automatically identifies and extracts evaluation criteria from natural language requests.
• Categorization: Distinguishes between two types of rubrics:
  • Hard Rules: Derived from explicit requirements (e.g., format, length, forbidden elements).
  • Principles: Abstracted, domain-agnostic quality criteria (e.g., clarity, correctness, reasoning).
• Comprehensive & Concise: Ensures rubrics cover all critical aspects of a request while remaining distinct and free of redundancy.
• Structured Output: Generates rubrics in a specific numbered list format, starting with "The response" and appending [Hard Rule] or [Principle].

Use Cases

This model is particularly well-suited for:

• Automated Evaluation: Generating criteria for automated assessment of LLM outputs.
• Reward Modeling: Creating synthetic rubrics to guide and improve the performance of large language models.
• LLM Alignment: Assisting in the alignment process by providing clear, structured feedback mechanisms.

For more technical details and to cite this work, please refer to the associated paper: OpenRubrics: Towards Scalable Synthetic Rubric Generation for Reward Modeling and LLM Alignment.