chenyifan0929/GenRubric-14B
GenRubric-14B is a 14 billion parameter model developed by chenyifan0929, built upon the Qwen3 Base architecture with a 32768 token context length. It specializes in query-specific rubric generation for scalable and fine-grained evaluation of large language models on open-ended tasks. This model improves rubric generation through a self-evolution mechanism, producing evaluation criteria that align well with expert-written rubrics and generalize across domains.
Loading preview...
GenRubric: Self-Evolving Rubric Generation for Scalable LLM Evaluation
GenRubric-14B is a 14 billion parameter model, part of the GenRubric family, specifically designed for query-specific rubric generation. Built upon the Qwen3 Base architecture, this model addresses the challenge of evaluating large language models (LLMs) on open-ended tasks by generating detailed, fine-grained evaluation rubrics.
Key Capabilities
- Automated Rubric Generation: Generates comprehensive, query-specific evaluation rubrics for LLM responses.
- Self-Evolution Mechanism: Improves rubric quality through a self-evolution process, reducing reliance on manual annotations.
- High Alignment: Produces rubrics whose induced evaluations show better alignment with expert-written rubrics.
- Domain Generalization: Demonstrates strong generalization capabilities to domains not encountered during training.
- Scalable LLM Evaluation: Facilitates scalable and fine-grained assessment of LLMs, particularly for complex, open-ended tasks.
Good For
- Automating LLM Evaluation: Researchers and developers needing to evaluate LLMs on diverse, open-ended tasks without extensive manual rubric creation.
- Ensuring Evaluation Consistency: Projects requiring consistent and objective evaluation criteria across various queries.
- Research in LLM Assessment: Studies focusing on improving the methodology and scalability of LLM evaluation.
- Reducing Human Annotation Effort: Scenarios where creating expert-written rubrics is time-consuming or impractical.