wassname comments on Sparse Autoencoders Work on Attention Layer Outputs

[wassname]

This make sense, as many adaptors use attn_out. It stands to reason that the layers, which make useful adaptors, also make useful probes.

Those exact layers are documented on the peft github.