Follow-up to #106, which pinned the huggingface downloads to 42dbf0ae (2025-06-13) so the existing 95M model keeps working.
Upstream has moved on. ctheodoris/Geneformer@main now offers Geneformer-V1-10M, Geneformer-V2-104M, Geneformer-V2-104M_CLcancer and Geneformer-V2-316M, and the dictionaries are *_gc104M.pkl. The gf-*-i4096 layout the component knows about is gone.
Migrating means:
- new
--model choices, and dropping the gf-<layers>-<dataset>-i<input_size> name parsing in src/methods/geneformer/script.py, since the V2 directories do not encode those the same way
- the
dictionaries_subfolder branch on 95M/30M needs a 104M arm
- dropping the revision pin once the component targets
main's layout
Worth deciding deliberately rather than drifting into it: V2 is a different model, so scores are not comparable with previous runs. Whoever picks this up should say in the PR whether the intent is to replace the 95M results or to report both.
Also relevant: #96, which proposed exactly this migration when the breakage was first reported.
Follow-up to #106, which pinned the huggingface downloads to
42dbf0ae(2025-06-13) so the existing 95M model keeps working.Upstream has moved on.
ctheodoris/Geneformer@mainnow offersGeneformer-V1-10M,Geneformer-V2-104M,Geneformer-V2-104M_CLcancerandGeneformer-V2-316M, and the dictionaries are*_gc104M.pkl. Thegf-*-i4096layout the component knows about is gone.Migrating means:
--modelchoices, and dropping thegf-<layers>-<dataset>-i<input_size>name parsing insrc/methods/geneformer/script.py, since the V2 directories do not encode those the same waydictionaries_subfolderbranch on95M/30Mneeds a104Marmmain's layoutWorth deciding deliberately rather than drifting into it: V2 is a different model, so scores are not comparable with previous runs. Whoever picks this up should say in the PR whether the intent is to replace the 95M results or to report both.
Also relevant: #96, which proposed exactly this migration when the breakage was first reported.