@capemox@raphaelsrty Ah, awesome. Are you planning to release embedding models based on these too? I was thinking it would be amazing to have the whole stack fully open and reproducible
@capemox@raphaelsrty Thanks for publishing the pretrained checkpoints! I'm also pretty resource constrained when playing with this stuff, partly by VRAM but mostly by lack of patience with longer jobs lol
I’m excited to join the speaker lineup at https://t.co/QxAGy0IKON Live! My session will explore Nuances of Binarized Embeddings-Based Retrieval
If you’re attending, let me know — would love to connect at the event.
My opinion on tokenmaxxing is companies shouldn’t mandate/constrain any tools at all and then evaluate software developers by output / (salary + token use)
bm25 is nice and all, but you won't believe how easy it is to improve upon it with and how much more you can squeeze from lexical features in @vespaengine
@tomaarsen Well, consider that part resolved 😉We "supported" it since long before the paper dropped, this could likely run just as well on years old Vespa version. Ofc would love to optimize for it specifically but that is likely gated on usage
Just added a sample app for how to search with hypencoder models on Vespa. A large meta-model that generates a small query-specific model that scores your docs - it feels like pure science fiction, but of course we can do it: https://t.co/3TAE6dnTvQ
@tomaarsen Of course that is purely theoretical unless we can show benefit from that thru training, more compelling results would be needed to raise interest