Working on my own variant of clap with my own text tokenization, planning to evaluate the ZS results against the CLAP created at microsoft. Was just the test set used for these metrics? Or are your zero shot results against the entire dataset?
Hi @ksasso1028 , we used only the testing sets to evaluate the downstream tasks. Note that some datasets have k-folds, so we considered all the folds as the testing set.
Working on my own variant of clap with my own text tokenization, planning to evaluate the ZS results against the CLAP created at microsoft. Was just the test set used for these metrics? Or are your zero shot results against the entire dataset?
Thanks.