DEV Community

Sarthak Agrawal
Sarthak Agrawal

Posted on

A perfect narrow score can still be the wrong general model

A distilled Qwen3-4B specialist improved the file-operations hard gate from 58% to 100%.

The same candidate reduced out-of-domain breadth from 59.6% to 42.3%.

Publishing only the first number would make the run look like a general model upgrade. It was not. The candidate became a routed specialist: useful when the task matches its narrow capability and unsafe to promote as the default planner.

This is why PostTrainLLM report cards include protected slices beside the target metric. A specialist should improve the behavior it was trained for, but the release decision also needs to show what it forgot. Routing is not an apology for the regression. It is the product boundary that makes the narrow gain usable.

The result also changed what "ship" meant inside the factory. A model can ship with a constraint when the evidence names that constraint and the runtime can enforce it. Without that boundary, the same artifact should be rejected as a general replacement.

The original measurements are documented at https://posttrainllm.com/devlog.

Top comments (0)