Show HN: Gatekeeper – Persuade a 34MB Jev-like model to let you into the castle

3 points by mazlix 19 days ago · 1 comment · 1 min read

Upfront caveat: it's English only! but I'm personally still just amazed that 34MB (fine-tuned BAAI/bge-small-en-v1.5 ) can fit so much knowledge that it's possible to have an iPhone 15 running these classifications entirely client-side on arbitrary english text in 10-30ms.

The game is a little trivial but thought it was a cool showcase, you type messages to the guard and they get classified as either friendly, bribe, threats, etc.

I've been playing a lot with Jev use-cases since it came out and then learned about fine-tuning my own models.

The way I went about it was just using Jev first via a proxy to record my calls and store them as my own training data to make running training easier. I built a whole oss platform around doing that with models (of various sizes) https://github.com/doppsh/dopp.

Would love to hear any feedback / discuss more on either that whole training process or the game itself.

Also I did get a ton of help from Opus doing all this, so some things might be a little rough around some edges but plenty of human time has been put in too. Very welcoming of any feedback/criticism