kev: a tiny Jev-like you can train on a MacBook

A Qwen2.5-0.5B-based decision model with block-causal masking so questions run in parallel and a pointer head that scores options; the HTTP API mirrors System One, so the official SDK works with only a base_url change.

Notes

0.799 held-out accuracy, ECE 0.031 after temperature scaling, ~160 ms for six-question requests, ~1h45m training on an Apple M5. Independent model.

source-backed — Public repository, docs or live artifact. About this label

Destinations

Filed under

• #replica
• #local
• #qwen
• #github

In collections

• Local & replica models →