RubricHub

Summary

RubricHub is a large-scale (approximately 110K), multi-domain dataset that provides high-quality rubric-based supervision for open-ended generation tasks. It is constructed via an automated coarse-to-fine rubric generation framework, which integrates principle-guided synthesis, multi-model aggregation, and difficulty evolution to produce comprehensive and highly discriminative evaluation criteria, overcoming the supervision ceiling of coarse or static rubrics. Leveraging RubricHub in a two-stage post-training pipeline (RuFT + RuRL) yields substantial gains in open-ended reasoning, enabling Qwen3-14B to achieve state-of-the-art performance of 69.3 on HealthBench, surpassing multiple proprietary frontier models.

Keywords

hf-dataset language-modeling · -reinforcement-learning · -qa parquet text datasets dask polars mlcroissant medical science wirting isntruction chat general has-paper

Topics

Language Modeling, Reinforcement Learning, QA

Research notes

• downloads=991; likes=270