Resolve the practical retrieval advantage of the larger embedding model

Determine whether the stronger final representation separation observed for the 560M Neurograf embedding model yields a reliable practical retrieval advantage over the 160M model under larger candidate pools, natural Romanian documents, graded relevance judgments, and hard negatives.

Background

The 560M embedding model eventually achieves better tracked end-state metrics than the 160M model, but the principal in-batch retrieval task uses only 48 candidates and approaches saturation for both models. This leaves little dynamic range for measuring meaningful differences between checkpoints.

The paper specifically identifies harder training and evaluation conditions—larger candidate pools, hard-negative mining, and retrieval against thousands or millions of passages—as necessary to determine whether the observed representation improvement translates into deployment-relevant retrieval performance.