News for hackers. Stuff that matters.
Re:Show HN: Jevstiller – Distill Jev into a local model, wit... by dotancohen on Tuesday September 29, 2026 @01:49PM (#49897433)
It would be great if we could correct Jev's incorrect answers, even on a separate endpoint. Let me tell it what Jev got wrong.

What type of head is that? What type of model is that head part of?

Re:Show HN: Jevstiller – Distill Jev into a local model, wit... by tgluck on Tuesday September 29, 2026 @03:11PM (#49898819)
Not today, but Interesting idea. The main motivation was a drop-in for an existing Jev setup, so the only teacher right now is Jev and the audit measures agreement with Jev. A correction would have to become a second label source that overrides Jev's for that input.

The head is a multinomial logistic regression: one linear layer plus softmax on top of a frozen sentence-embedding model (bge-small by default, swappable). That head is the entire local model, the encoder is off the shelf and never changes.

Re:Show HN: Jevstiller – Distill Jev into a local model, wit... by dotancohen on Tuesday September 29, 2026 @06:59PM (#49901981)
Yeah, I kinda figured that head was the whole model, the way you phrased it. scikit-learn?
Re:Show HN: Jevstiller – Distill Jev into a local model, wit... by tgluck on Tuesday September 29, 2026 @07:08PM (#49902072)
No, plain numpy. It's full-batch Adam on cross-entropy against soft targets, about 80 lines.
Re:Show HN: Jevstiller – Distill Jev into a local model, wit... by dotancohen on Wednesday September 30, 2026 @04:26AM (#49905988)
I'll look into Adam, thank you!