ProSST / README.md
tyang816's picture
Add ProSST structure tokenizer files for rem2 / VenusREM2
0941c17 verified
|
Raw History Blame Contribute Delete
773 Bytes
metadata
license: apache-2.0
tags:
  - protein
  - structure
  - ProSST
  - rem2
  - VenusREM2

ProSST structure tokenizer (rem2 / VenusREM2)

Structure-token assets used by rem2 when you run --model prosst or --model venusrem2.

On first use, rem2 downloads only the files it needs into ~/.cache/rem2/weights/prosst/static/:

File Used by
static/AE.pt All ProSST vocab sizes
static/{K}.joblib --model prosst-{K} and VenusREM2 members. K ∈ {20, 64, 128, 512, 1024, 2048, 4096}

VenusREM2's official ensemble uses K ∈ {20, 128, 512, 1024, 2048, 4096} (not 64).

These files are the local structure tokenizer (AE.pt + k-means vocab). They are not the ProSST language-model weights.