/alignmentresearch/ Eliciting Latent Predictions from Transformers with the Tuned Lens

Code Link

https://github.com/alignmentresearch/tuned-lens

Description

We analyze transformers from the perspective of iterative inference, seeking to understand how model predictions are refined layer by layer. Code: https://github.com/alignmentresearch/tuned-lens

Retrieved

2023/03/16

Stars

TOP