Edu-QuRating
A research paper published on arXiv that proposes a multi-dimensional method for scoring and filtering educational data. It constructs reusable educational quality scorers by distilling pairwise preference judgments, serving as an alternative to conventional single-scalar educational value filtering.
Timeline (last 90 days)
Edu-QuRating specializes the evaluation dimensions into six criteria for the educational domain, covering accuracy, structure, engagement, and others.
The research team used a scorer to score 322,250,000 FineWeb-Edu-Fortified documents.
The model trained using Edu-QuRating-filtered mixed sets achieved a higher overall observed accuracy than the FineWeb-Edu baseline model across nine benchmark tests, with performance gains concentrated on specific tasks.
The researchers used Edu-QuRater scores as a reward term for GRPO post-training, with Qwen3-4B as the base model.
The best-performing Edu-QuRater achieved an average accuracy of 0.917 in reproducing GPT-4.1-mini pairwise judgments on the held-out set.
The research team used the scorer to score 322,250,000 FineWeb-Edu-Fortified documents and constructed a curated pre-training data mixture.
The paper "Edu-QuRating: Multi-Dimensional Educational Data Curation with Distilled Pairwise Judgements" proposes a pipeline for multi-dimensional educational data scoring and curation.
The researchers used Edu-QuRater scores as a reward term in GRPO post-training, with Qwen3-4B as the base model in the experiments.
After combining the Edu-QuRater reward with the answer structure reward, the generated responses outperformed the Qwen3-4B base model on both the instructional quality and instruction-following dimensions.
In controlled single-run pre-training comparison experiments, the model trained on the Edu-QuRating-curated mixture achieved higher overall observed accuracy across nine benchmarks than the FineWeb-Edu baseline model, with performance gains concentrated on specific tasks.
3 more timeline events
All Facts (13)
Edu-QuRating specializes the evaluation dimensions into six criteria for the educational domain, covering accuracy, structure, engagement, and others.
50%UnverifiedThe research team used a scorer to score 322,250,000 FineWeb-Edu-Fortified documents.
50%UnverifiedThe model trained using Edu-QuRating-filtered mixed sets achieved a higher overall observed accuracy than the FineWeb-Edu baseline model across nine benchmark tests, with performance gains concentrated on specific tasks.
50%UnverifiedThe researchers used Edu-QuRater scores as a reward term for GRPO post-training, with Qwen3-4B as the base model.
50%UnverifiedThe best-performing Edu-QuRater achieved an average accuracy of 0.917 in reproducing GPT-4.1-mini pairwise judgments on the held-out set.
50%UnverifiedThe research team used the scorer to score 322,250,000 FineWeb-Edu-Fortified documents and constructed a curated pre-training data mixture.
50%UnverifiedThe paper "Edu-QuRating: Multi-Dimensional Educational Data Curation with Distilled Pairwise Judgements" proposes a pipeline for multi-dimensional educational data scoring and curation.
50%UnverifiedThe researchers used Edu-QuRater scores as a reward term in GRPO post-training, with Qwen3-4B as the base model in the experiments.
50%UnverifiedAfter combining the Edu-QuRater reward with the answer structure reward, the generated responses outperformed the Qwen3-4B base model on both the instructional quality and instruction-following dimensions.
50%UnverifiedIn controlled single-run pre-training comparison experiments, the model trained on the Edu-QuRating-curated mixture achieved higher overall observed accuracy across nine benchmarks than the FineWeb-Edu baseline model, with performance gains concentrated on specific tasks.
50%UnverifiedEdu-QuRating specializes QuRating's general text quality evaluation dimensions into six criteria specific to the educational domain.
50%UnverifiedEdu-QuRating uses GPT-4.1-mini as the judge to make pairwise preference judgments on document pairs.
50%UnverifiedEdu-QuRating distills pairwise preference judgments into lightweight sequence classification models, training reusable Edu-QuRater scorers.
50%