Turn detection experiment: Jev hits 16/16 while tiny-BERT Laya flunks on own data

tobowers · x · 2026-09-22

The author ran a turn-detection experiment comparing Laya (a Jev-like architecture built on a tiny BERT) against Jev itself. Jev scored a perfect 16/16, while Laya hit roughly 9-10/16 — near random chance — and sometimes broke completely depending on how the state was worded. The takeaway: Laya is a solid architecture, but beware of benchmark scores and always test on your own datasets.

Original post →

More from coding & agent

coding & agent channel →