MolSAEinterpretability lab
frozen base · GP-MoLFormer-Uniq · 12L·768d · SAE L8 batch-topk

Input molecule

SMILES

Per-token activation

layer 8
Tokens are from the model tokenizer. Color intensity reflects the strongest feature score for that token. You can select multiple tokens; mapped atoms are highlighted in 2D/3D.
low
high
loading structure...
Aspirinresolved name
formula
0heavy atoms
0tokens
top feature

Top activating features

0 shown
No features available for this molecule.
loading...
name Caffeineactive 0
loading...
name Nicotineactive 0shared 0

SMILES set

5 molecules
Builds a molecule × feature activation matrix and computes feature co-occurrence across the set. Use format: SMILES Name to set a custom molecule name.

Molecules

Activation matrix
molecules (rows) × most-shared features (cols)
Analyzing set...

Feature co-occurrence — pairs that fire together

Not enough molecules for pair statistics yet.