# Jevlike
Jevlike is an open-source starter model by vinnylarouge that "chooses among a changing list of text options". It's inspired by [[Jev]], but it's closer to a trainable classifier than to a general decision model.
How it works: each option becomes a query vector that attends over the context tokens. That gives one context vector per option, which is scored with a shared dot product; a softmax turns the scores into probabilities. Everything happens in one pass, and it's light enough to train from scratch on a CPU.
Results from the repo:
- ~98% accuracy on synthetic menus
- 26% on Wikispeedia link choices with a frozen Qwen2.5-0.5B encoder (8% for the control); 29% with a byte encoder trained from scratch on 40,000 clicks
- About 100x faster than making a small decoder write 400 tokens
The catch: you need your own labels and a training run. Jev's appeal is answering new questions with zero training; jevlike doesn't do that. The repo includes fun examples though (a Doom player and a chess controller using its vision scorer), and it's a good codebase to learn how option scoring works under the hood.
## References
- [jevlike on GitHub](https://github.com/vinnylarouge/jevlike)
- [Open-source Jev alternatives (Apidog)](https://apidog.com/blog/openjev-open-source-jev-alternatives/)
## Related
- [[Jev]]
- [[System One Models]]
- [[SemIf]]
- [[Kev]]