kev trains a typed-decision model on a MacBook in 1h45m and lands within 6-7 accuracy points of the commercial reference
jaredpalmer/kev (created 2026-09-17, 733 stars, pushed 09-20) is a LoRA adapter plus a custom readout head on Qwen that answers many typed questions about one document in a single forward pass, across yes/no, multiple-choice and ordered-scale question types. Training kev-0.5b takes about 1 hour 45 minutes on an Apple M5, larger variants 40-70 minutes on one H100, using ~1,000 records each from 10-13 public sources (Banking77, AG News, MNLI, BoolQ, SST-5, Yelp) plus synthetic policy pairs. On held-out out-of-domain sources kev-4b scores 0.790 accuracy / 0.328 Brier versus the Jev reference at 0.857 / 0.211, serves in about a second on a 32GB Mac in bf16, and speaks the official typesafe-sdk wire format so you can point an existing client at localhost.
↳ Follow the thread