canrager Goto Github PK

algorithms

Coding exercises in algorithms and data structures

api-connector

articulate_rules

Investigating the ability of LLMs to articulate rules for classification tasks they can solve.

attribution-patching

Method for identifying circuits in transformer models

brackets_interpretability

chess_llm_interpretability

Visualizing the internal board state of a GPT trained on chess PGN strings, and performing interventions on its internal board state and representation of player Elo.

clas

Circuit-Level Activation Steering

cleanup_nodes

dictionary_learning

Precompute model activations for SAE training and evaluation

eap

Edge Attribution Patching for TransformerLens

expert-sae

Examining the role of experts in MoE models using dictionary learning.

feature-annotation

Web-app for crowdsourced annotation of SAE features

feature-clustering-webapp

Unsupervised method for clustering sentences with diverse contexts into groups where similar features are involved for next token prediction.

indirect_object_identification

Replicating the results of the paper "Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small" by Wang et al.

mats_sae_training

Training Sparse Autoencoders on Language Models

nanogpt

Rebuilding a minimalistic version of GPT along Andrej Karpathy's lecture

nnsight_workshop

Playground for manipulating activations in large language models with the nnsight module by BauLabs.

ravel

Evaluate interpretability methods on localizing and disentangling concepts in LLMs.

representation-learning

Learning State Representations for Quantum Systems

sae_eval

shift_eval

sparse-coding

Improvements on dictionary learning for language model interpretability: (1) training and (2) evaluation of sparse autoencoders.

superposition_replication

Replicating results from "Toy models of Superposition" by Anthropic