Building a Codebase Classifier with Jev
A developer shares building a codebase classifier with Jev, suggesting it may solve overengineered code from agents, and asks what to test next.
Jev model performs well on BANKING77 benchmark
Using the same pinned Jev version on 3,080 official BANKING77 test cases, accuracy reaches 80.75% without examples and 91.36% with all 2,446 labeled examples in context — 327 more correct, no weight updates, and only $10.92 of input tokens.
BANKING77, same pinned Jev version, same 3,080 official test cases: 80.75% with no examples, 91.36% with all 2,446 labeled examples in context. 327 more correct, no weight updates, $10.92 of input tokens.
· 0 likesOpen on X
More like this
Jev Enables LLM-as-a-Judge at 20–200x Speed and 40–400x Lower Cost
Jev was used to pick which model is AGI, offering 20–200x speed and 40–400x lower cost, making LLM-as-a-judge extremely fast and nearly free.
Testing apps with TypeSafe Jev & OpenCode
Filip shares a small demo showing how fast testing apps can be with TypeSafe Jev and OpenCode.
