Lumo that learns each user, with no logs kept Hi Proton team, I'm Troy, I run a small lab in Spain (Garnet Taurus Ltd) working on mod
Hi Proton team, I'm Troy, I run a small lab in Spain (Garnet Taurus Ltd) working on models that keep learning into their own weights after release.
Lumo can't personalise from stored chats because it doesn't keep logs. Our approach could fill that gap: what a user chooses to teach Lumo is learnt into a small add-on file (about 10 MB) that belongs to that user, instead of being stored as text. Deleting the file deletes what it learnt, and the base model is never changed.
We've already tested this on models Lumo serves. On Apertus-8B, one day of new facts learnt into an add-on went from 11.7% to 91.0% correct, with general skills unchanged. On OLMo 3 7B it went from 6.8% to 88.7%. This week, on a 1.2B on-device model, a separate add-on for one person reached 87 to 90% on questions it had never seen, and when it answered with confidence it was wrong on 2 of 76.
These are early results on test data. I'd be glad to run it on Apertus 1.5 with data you choose. Would this be interesting for Lumo?
Troy Clifford