pixelart34·2 年前·議論Exactly, love the openness about the training data which none of the big labs disclose, the v2 apparently uses a better mix according to the model card and scores