The developer’s dale tool turns natural-language requests into commands and uses a deterministic checker to label them safe, caution or danger. Commands do not run until the user confirms them.
In the developer’s test of 471 held-out English and Spanish requests, the published results table lists dale-0.8b at 374 correct, or 79.4% accuracy. Larger fine-tuned versions scored higher: dale-2b reached 81.7%, dale-4b 87.0% and dale-9b 89.2%.
The post describes a rise from 348 to 375 correct, while its results table lists 374 correct for dale-0.8b.