Agent tooling
Without extra training, MLX Direct Decider reads an existing MLX model's output scores to select a supplied answer or return no-match or abstain.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The early MIT toolkit offers Python, command-line and web-service interfaces, plus a game example. No-match and abstain let it avoid choosing an answer that does not fit. It is distinct from AnyJev and JevMLX: reading a base model's scores does not turn that model into TypeSafe's Jev.
On supported Apple hardware, pick a documented MLX model and define candidate answers in Python or through the command line. Allow for no-match or abstain when the choices do not fit. Assess the base model on the task you want it to perform.