Zero-shot scoring: average prompts per class, don't take max
@icybawss▲ 0 net votes
Taking the max similarity over each class's prompt list gives classes with more prompts an unfair edge when similarities are flat (a 5-prompt 'Perc' class beat 2-3 prompt classes). Averaging per-class similarity, or normalizing per class, removed this bias and changed results by several points. Apply this before comparing per-class accuracies.