google

/

gemma-scope-2b-pt-res

like 4

SAEs on the raw embedding vectors (not including the sqrt(d_model) scaling that happens inside the model fwd pass).
Not trained on BOS, PAD, EOS and any token that didn't occur in the training data.