[bick_nyers]

I wonder if you would want to use an earlier layer as opposed to the penultimate layer, I would imagine that the LLM uses that layer to "prepare" for the final dimensionality reduction to clean the signal such that it scores well on the loss function.