Evaluate a 64-dimensional ItColBERT variant and establish an Italian MTEB-style evaluation

Evaluate a 64-dimensional variant of ItColBERT to determine whether halving the late-interaction index size preserves useful retrieval quality, and establish an MTEB-style community evaluation for Italian retrieval in the absence of an existing Italian MTEB.

Background

ItColBERT uses 128-dimensional token vectors, and the paper identifies index size as a practical concern for multi-vector retrieval. A 64-dimensional version could therefore reduce storage requirements, but its retrieval-quality trade-offs were not studied.

The paper also evaluates Italian retrieval using four existing benchmarks, several of which are translated or partially in-domain. Because no Italian MTEB exists, a broader community evaluation remains unavailable for systematically comparing Italian retrieval systems across tasks and domains.