…but it works we try a bunch of baselines like selecting data based on likelihood or even the top-k per-class our method (SELECT) can train a model to similar performance, with access only to model weights and some web data

→ View original post on X — @jxmnop