This is a tied-crosscoder of 16384*3 latents trained on the 14th layers of Gemma 2 2b and Gemma 2 2b it. We trained for 250M tokens on a mixed dataset of the pile and LMSys chat data.
This is a tied-crosscoder of 16384*3 latents trained on the 14th layers of Gemma 2 2b and Gemma 2 2b it. We trained for 250M tokens on a mixed dataset of the pile and LMSys chat data.