Clarification about the DeepSEA fine-tuned checkpoint (gena_web_deepsea)

[ziangshi]

Hi, thank you for releasing GENA-LM — it’s an excellent resource.

I have a question regarding the DeepSEA downstream model. On HuggingFace there is a branch:

https://huggingface.co/AIRI-Institute/gena-lm-bert-base-t2t/tree/gena_web_deepsea

This appears to be the model used for the GENA-Web demo. Could you please clarify the following:

Is the gena_web_deepsea checkpoint the same model used in the GENA-LM paper’s DeepSEA downstream evaluation (Table 3)?

If not, is the paper’s fine-tuned DeepSEA checkpoint available anywhere?

If no pretrained checkpoint is available, could you provide the main hyperparameters you used to fine-tune DeepSEA starting from gena-lm-bert-base-t2t?

The example code in the downstream_tasks/DeepSea directory indicates that the model seems fine-tuned from an existing checkpoint. Would you please also clarify the process and which checkpoint was used as the initialization for the DeepSEA fine-tuning in your experiments?

I hope to get a clear view of the correct preprocessing steps and fine-tuning configuration from you.

Thank you!