Can I read Dictionary Learning in Language Models on EtoBox?

Dictionary Learning in Language Models by dyangcad600 is a document available to read on EtoBox.

What is Dictionary Learning in Language Models about?

The document discusses advancements in interpretable dictionary learning for language models (LMs) using board game transcripts, specifically chess and Othello, to evaluate the quality of sparse autoencoders (SAEs). It introduces novel metrics for assessing SAE performance and a new training technique called p-annealing, which enhances SAE capabilities. The authors present their findings on how well SAEs can capture interpretable features of game states and provide open-source access to their trained models

Author

dyangcad600

Language

EN