Edu-QuRating introduces a multi-dimensional approach to evaluating educational data, moving beyond single scalar properties. The system utilizes an LLM judge to assess document pairs based on six educational criteria, including accuracy, engagement, structure, and appropriateness. These pairwise judgements are then distilled into reusable Edu-QuRaters, which can score individual text chunks. Initial experiments demonstrated that a single Edu-QuRater recovered held-out GPT-4.1-mini pairwise judgements with a mean accuracy of 0.917.
Applied to corpus filtering, Edu-QuRaters were used to score 322.25M FineWeb-Edu-Fortified documents. Subsequent pre-training with the resulting mixture yielded higher aggregate accuracy across nine benchmarks compared to a baseline. These gains were most pronounced in specific tasks.
Furthermore, Edu-QuRater scores were integrated as reward terms within a GRPO post-training framework. Combining these rewards with answer-structure rewards resulted in responses preferred to the Qwen3-4B base model regarding both pedagogical quality and instruction following.