Trouble recreating zero shot results on protein gym

[morgankohler]

Hello thanks for the work. I am running the compute_fitness script on protein gym and am having a hard time getting around the stated spearman of 0.517. I am running the script unchanged with the a2m EVCouplings alignments from huggingface, the struc sequences from the tar file in the repo, and a sample ratio of 0.8. However the average correlation is only 0.305. Any ideas why I am not getting a higher result?

[tyang816]

You may have made some modifications to the code, perhaps using Claude for diff comparison?