[Legend2440]

I don’t think benchmark overfitting is as common as people think. Benchmark scores are highly correlated with the subjective “intelligence” of the model. So is pretraining loss.

The only exception I can think of is models trained on synthetic data like Phi.