Comment by Legend2440

Comment by Legend2440 16 hours ago

0 replies

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.