AgenticDS preprint reports better training-data search and downstream accuracy

Researchers report that AgenticDS, a framework for finding model training datasets, improved search ranking quality and downstream training accuracy in a new preprint. On HuDS-Bench, its NDCG@10 ranking score improved by 25.7% over the strongest dense-sparse hybrid search system and by 18.9% over a commercial language-model agent. AgenticDS converts search requests into structured plans, executes constrained retrieval and graph searches, and checks supporting evidence before accepting rankings. The benchmark contains 200 queries and 24,908 graded judgments pooled over 15 systems. A controlled training study found that its retrieved datasets delivered the best accuracy in mathematics, code and medicine; the abstract gives no accuracy margins.