← Glossary Term
IFEval
A benchmark that tests whether a model follows clear, checkable instructions.
IFEval gives a model instructions that can be checked automatically, such as “write in under 100 words” or “include the word blue twice.” The score is the share of instructions it follows.
It is a narrow test of obedience to format and rules, not of how smart or helpful the model is overall.
Mentioned in