Hacker News Reader

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

No, plain numpy. It's full-batch Adam on cross-entropy against soft targets, about 80 lines.

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Obviously it only helps when the same question is asked many times, but that's the case it's built for, and the case I have, every Jev question in my other project repeats thousands of times, and most Jev uses I've seen online look the same

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

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.

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Thanks.

Drop-in yes: point TYPESAFE_BASE_URL at it and nothing else changes.

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Not really

they distill different things. model2vec distills a sentence transformer into static embeddings, so the output is a faster general-purpose encoder.

Jevstiller keeps the encoder frozen (bge-small by default) and distills Jev's decisions on one specific question into a small head on top of it

tgluckon Show HN: Jevstiller – Distill Jev into a local model, with a disagreement bound

Author here. This puts a proxy in front of repeated Jev classification calls. At first everything goes to Jev; from Jev's answers it trains a small head on frozen sentence embeddings, picks a confidence threshold with an exact finite-sample bound so that at most 2% of all requests get an answer Jev wouldn't have given, and then answers the confident share locally at ~15 ms on a CPU. A permanent 2% audit keeps checking; if agreement breaks, everything falls back to Jev and it retrains.

Known limits: agreement is not accuracy (if Jev is wrong, so is the local model); coverage tracks how consistent Jev itself is (22% on noisy tweet tasks, 80% on news); it speaks Jev's API only, an OpenAI-compatible front is on the roadmap. Since 0.4.0 the guarantee can also cover "would Jev have been unsure", which matters if your code routes low-confidence answers to review. Apache 2.0.