Bug description
When using ESM-IF1 to generate many sequences the occasional sequence (approximately 1/10,000) will contain erroneous characters such as 'af2' or 'eos'. The rest of the sequences seem to be valid sequences and fold as expected when checked with alphaFold2.
Reproduction steps
Running model.sample(coords, temperature=temp) thousands of times
Bug description When using ESM-IF1 to generate many sequences the occasional sequence (approximately 1/10,000) will contain erroneous characters such as 'af2' or 'eos'. The rest of the sequences seem to be valid sequences and fold as expected when checked with alphaFold2.
Reproduction steps Running model.sample(coords, temperature=temp) thousands of times
Expected behavior Generate normal sequences
Logs NA
Additional context