Closed linhan12 closed 6 days ago
Hi! the "_recover_indices.npy" was inserted to shuffle the target dataset and recover its order during embedding extraction since the difference of language length for each task caused different speed for each GPU. We removed it due to its redundancy, but it seems that it caused the bug. I recovered the original extract_embed.py file. Please check!
hi, I want to ask where is "tan_act_37111519_msa_recover_indices.npy" as you referred in line 24 in COINCIDE_code/COINCIDE_cluster/tinyllava/eval/score/merge_values.py? or how can we generate this .npy file?