Are the downloadable finetuned weights for secondary structure prediction intra- or inter-familty trained?

[shanry]

I noticed that the the split in the archiveII dataset (fam-fold) is intra-family based. However, the paper claims that the fine-tuning on archiveII used an inter-family split.

Could the author clarify a bit about how the downloadable weights are fine-tuned?

[RJPenic]

Hello 😄,
As stated in the paper: "The dataset of 3865 RNAs from nine families was split nine times, and in each split, a different family was held out for evaluation while the other eight families were used for training and validation."

I just tried out weights that were used for SRP family evaluation and results are as expected. Are you sure you are using the right weights? File rinalmo_giga_ss_archiveII-srp_ft.pt contains weights of the model trained (and validated) with all families except SRP.