Writing C++ with Jev by picking valid next tokens
The author experimented with TypeSafe's Jev decision/classifier model for C++: at each step it was given only grammatically valid next tokens to pick from, avoiding free-form generation.
First Benchmark Run: LFM3-VL-2B on JevBench
John Wong shares LFM3-VL-2B's JevBench results after 200 LoRA steps with a frozen backbone: dev 0.657 / test 0.452, noting it's not comparable to the closed-source 0.996, currently ~10 points behind the best open-source model of similar size, and 25–30 points behind Jev itself. Further optimization is planned.
第一次跑分，先记个数。 LFM3-VL-2B 冻结主体 + 200 步 LoRA，JevBench 路线 0.657 dev / 0.452 test，原始准确率，跟闭源那 0.996 不是一把尺子。 同量级开源最好 0.753，我们差 10 个点，落在 2B 中下段；离 Jev 本体差一档，25 到 30 个点。 后面接着优化，数会动。
· 0 likesOpen on X
More like this
Using Jev for Code Review
The author thinks code review is a really good use case for using Jev.
Fascin: A Jev-Powered PR Triage Tool
A developer shares Fascin, a PR triage tool built with Jev that assesses urgency, change type, and merge readiness, noting Jev handles the actual reasoning.
