Yah, that’s a fair point. I had assumed it’d remain relatively similar given that the training data would be smaller for languages like Haskell versus languages like Python & JavaScript.