14.5 hours on one RTX 3090: finetuned multi-vector retriever beats all general-purpose models
tomaarsen · x · 2026-08-26
Tom Aarsen, author of sentence-transformers, published an in-depth tutorial: v6.0 introduces a fourth model type, MultiVectorEncoder, for ColBERT-style late interaction retrieval with a complete training pipeline. Everything runs via pip install -U "sentence-transformers[train]".
The post walks through each component of finetuning multi-vector models — model, datasets, losses, training arguments, evaluators, and the trainer — with practical examples.
For evaluation, his multi-vector-encoder/mLateOn-medical, trained in 14.5 hours on a single RTX 3090, easily outperforms every general-purpose retrieval model he could find on his medical retrieval benchmark: dense, sparse, lexical, and multi-vector alike. The method also trains strong new multi-vector models from scratch.
Related event: Fine-tuning ColBERT on a single RTX 3090 beats general-purpose retrievers(14 posts)→
More from coding & agent
- raylib 6.2 ships with 300+ improvements and 100 new contributors, 20 new API functions — jsuarez · 2026-10-16
- "Vibecoding" is as misleading a label as "AI" and "SI," dev complains — eschadiol · 2026-10-16
- Doc2Harness: New benchmark finds best coding agent passes only 53% across 329 cases — _akhaliq · 2026-10-16
- Perplexity open-sources SPACE, the Firecracker microVM sandbox behind Perplexity Computer — perplexity_ai · 2026-10-16
- Blackstone goes all-in on Factory for autonomous software development — matanSF · 2026-10-16
- Longtime Windsurf user cancels $10/month legacy plan, citing AI subscription fatigue — tristanbob · 2026-10-16