living-models

/

Botanic1-S

Like 2

Follow

Living Models 25

Hi Living Models team,

Thank you for releasing Botanic1-S and the associated code/resources.

I have a question about the relationship between the currently released living-models/Botanic1-S checkpoint and the long-context Botanic1-S model described in the BOTANIC-1 preprint.

From the model card, my understanding is that the currently released Botanic1-S was pretrained with an 8,192-token sequence length. However, the paper also describes a context-extension curriculum for a Botanic1-S-sized model:

8,192 → 16,384 → 32,768 → 65,536 → 131,072 bp

and refers to a final long-context model/checkpoint, which I understand as the Botanic1-S-128k model used in the long-context experiments.

Could you please clarify the following?

1. Is the currently released living-models/Botanic1-S checkpoint the 8,192-bp pretrained model, or does it already include the context-extension training up to 131,072 bp?

2. If the currently released checkpoint is the 8k version, is the final 131,072-bp context-extended Botanic1-S checkpoint publicly available somewhere?

3. If it is not yet public, are there plans to release the 128k checkpoint?

4. For sequence lengths substantially longer than 8,192 bp, such as ~78 kb, would you recommend using only the context-extended checkpoint rather than directly running the current 8k checkpoint at a longer sequence length?

My use case is nucleotide-resolution functional track prediction from ~77,824-bp plant genomic sequences, so the distinction between an 8k-pretrained checkpoint and a genuinely context-extended 128k checkpoint is very important for us.

Thanks very much for the clarification and for making these resources available.

Best regards,
Jianhong Liu