Our summary
Fastino’s GLiNER2.5-Decide answers questions with supplied choices and can check related answers together. Mary Newhauser describes a model small enough to run on your own computer without a graphics card. It is a competing model, not Jev; the article compares several alternatives using Fastino’s own test collection.
Suppose one question detects a harmful instruction while another labels the same message safe. You can supply a rule that detected harm requires an unsafe answer. The model looks for the highest-scoring combination allowed by that rule, then reports whether the combination satisfies the rules. Consistency does not prove the judgments correct.
The reported accuracy and speed depend on the tests, computers and text lengths used. One comparison model, JevK5, was made separately to reproduce Jev’s approach; it is not TypeSafe’s own model. Fastino used a different test collection. JevMade has not repeated the tests or verified running the model without an internet connection.
Key takeaways
- Write explicit rules when several answers must agree; checking them together differs from asking independent questions.
- Answers can obey every rule and still be wrong; test against examples whose correct answers are known.
- Reported results depend on which model, tests, computer and text lengths were used, not just the model’s name.