but maybe im being pedantic, it’s still deep nns + foundation model finetuning so definitely not “classical machine learning” to your point :)

i don’t think that’s true. it’s a graph neural network not transformers and it’s finetuned from a 40m parameters model so definitely trainable on older gen gpus

I doubt it’s intentional upweighting but it was likely preferred by the median rater awhile ago and it’s baked into the reward model data.

that’s right