Here's what was shared previously to address this issue:

the dev_meta table and the dev/validation dataset contain 3227 samples/audios, the test_meta/test dataset that will be uploaded later contains 5064 samples, while the submission file should contain all IDs (8091). The easiest solution is to concatenate your dev_meta predictions with, a " " space for the test_meta transcripts.

Let's take your LB WER as an example:

wer_all=0.492613365 is the WER for all submission ids (8091), the WER for spaces is 1 (for 5064 test samples)

so to calculate the WER for only the dev_metadata predictions, it will be :

wer_dev = 3227*wer_all/(8091-5061*wer_all) = 3227/(8091/0.4926 - 5064/1) = 0.284

your model is not that overfitted hh