Shosuke Suzuki

@jwohlwend That's the key question, thank you for asking! You are correct, the confidence model alone **isn't a reliable way** to pick out the best structure from the samples. For...

Great! I'll add the documentation and request your review once it's updated.

@jwohlwend Hi, I've updated the `prediction.md` - Added the new arguments - Included the note about the limitations of confidence scores - Made a cosmetic tweak to the table's Markdown...

Hi @jwohlwend, This PR is now updated and ready for review. - The `--mask_msa` logic was adjusted to apply masking **before one-hot encoding**, ensuring consistency with how MSA masking operates...

@sugardoll223 If you want to try multi-GPU distributed training, you can use [Kotomamba](https://github.com/kotoba-tech/kotomamba), as the Mamba2 block likely requires a single GPU.

At the time I opened this, using a shared `torch.Generator` felt like a useful improvement for reproducibility. Since then, I've realized that `seed_everything` from PyTorch Lightning already provides a good...