8000 alternate implementation without transformerlens · Issue #29 · FailSpy/abliterator · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

alternate implementation without transformerlens #29

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
Open
wassname opened this issue Mar 11, 2025 · 1 comment
Open

alternate implementation without transformerlens #29

wassname opened this issue Mar 11, 2025 · 1 comment

Comments

@wassname
Copy link

I remember when we were working on this together the paper first came out. Anyway you might be interested in my implementation using baukit instead of transformerlens. The advantage is that you don't need to map the weights, just the layer names. It makes exporting the model easier.

https://github.com/wassname/abliterator
@wassname
Copy link
Author
wassname commented Mar 11, 2025 •
edited
Loading

Also, the bottleneck here is often storing large activations in mem. Here's a nice way to cache them to disk, so that you can abliterate with larger datasets: https://github.com/wassname/activation_store

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
1 participant
@wassname
0