sparse-autoencoder-training
orchestra-research892 installs
Find interpretable features in language models.
What it does
Guides training and analysis of sparse autoencoders so users can identify interpretable features in language-model activations.
When to use it
Use when discovering interpretable model features, studying superposition, steering or ablating features, or analyzing safety-relevant features.
How to use it
It changes how the agent implements SAE training and analysis by prescribing configuration choices and validation checks.
What you provide
- An existing project
Uses
SA
Latest review
What you need · 4
Requires Python 3.10 or later.
Declares sae-lens version 6.0.0 or later as a dependency and provides its installation command.
Access · 6
This skill
W&B
Write
Logs training metrics to W&B
gpt2-small
Read
Loads pretrained language-model weights