andyrdt

andyrdt.github.io

ARENA_3.0

circuit-breakers

Improving Alignment and Robustness with Circuit Breakers

CircuitsVis

Mechanistic Interpretability Visualizations using React

eleutherai_sae

Sparse autoencoders

iclr

llm-attacks

mats_sae_training

mi

Repo to track miscellaneous mi stuff

refusal_direction

Code and results accompanying the paper "Refusal in Language Models Is Mediated by a Single Direction".

andyrdt/refusal_direction

Code and results accompanying the paper "Refusal in Language Models Is Mediated by a Single Direction".

andyrdt/eleutherai_sae

Sparse autoencoders

andyrdt/mats_sae_training

andyrdt/andyrdt.github.io

andyrdt/ARENA_3.0

andyrdt/circuit-breakers

Improving Alignment and Robustness with Circuit Breakers

andyrdt/CircuitsVis

Mechanistic Interpretability Visualizations using React

andyrdt/iclr

andyrdt/llm-attacks

andyrdt/mi

Repo to track miscellaneous mi stuff

andyrdt/path_patching

Implementation of path patching & activation patching (will eventually add to TransformerLens).

andyrdt/SycophancySteering

Modulating sycophancy in llama-2 via activation steering

andyrdt/TransformerLens

A library for mechanistic interpretability of GPT-style language models

andyrdt/wmdp

WMDP is a LLM proxy benchmark for hazardous knowledge in bio, cyber, and chemical security. We also release code for RMU, an unlearning method which reduces LLM performance on WMDP while retaining general capabilities.