New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Proposal] Support Anthropic's JumpReLU improvements #427

Open

chanind opened this issue Feb 11, 2025 · 0 comments

Collaborator

chanind commented Feb 11, 2025

In https://transformer-circuits.pub/2025/january-update/index.html, Anthropic documents improvements to JumpReLU SAE training. These include the following:

Using Tanh for the straight-through estimator
Adding a loss to discourage dead features
Allowing gradients from the straight-through estimator to flow through to model parameters

We should support these options for JumpReLU training in SAELens

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment