Which model would you like to support?
https://huggingface.co/lightonai/LateOn
What are the main advantages of this model?
Beats every existing ColBERT model, including those 4× its size (Jina ColBERT v2 at 559M, Arctic Embed L v2 at 568M).
Holds up under decontamination: when training-overlap samples are stripped from the BEIR corpora, LateOn climbs to 60.36 nDCG@10 on the 12-dataset decontaminated split — first place overall.
Uses fully open data for both pre-training and fine-tuning, with all signals released as metadata so you can rebuild, extend, or replace any filter.
https://x.com/antoine_chaffin/status/2060845436378783970?s=20
Which model would you like to support?
https://huggingface.co/lightonai/LateOn
What are the main advantages of this model?
https://x.com/antoine_chaffin/status/2060845436378783970?s=20