Skip to content

Conversation

@jordansauce
Copy link
Collaborator

Description

See issue #54: Currently our feature dashboards rely on an old commit of Callum McDougall's sae_vis library. Specifically, callummcdougall/sae_vis@b28a0f7.
This PR will make our code compatible with the latest version of sae_vis (0.2), which was significantly restructured in March 2024. The recent changes include:

  • The ability to view multiple features on the same page (rather than just one feature at a time)
  • D3-backed visualisations (which can do things like add lines to histograms as you hover over tokens)
  • SAE-class agnostic functions parse_feature_data() and parse_prompt_data() are now added to sae_vis, so we can remove them from our codebase - see Added SAE class agnostic functions callummcdougall/sae_vis#17
  • Significantly modified data storing classes and HTML generating functions

No more major restructures of sae_vis are planned from now (callummcdougall/sae_vis#15 (comment)) so this is a good time to upgrade.

However it also seems much slower to generate prompt-centric dashboards now (either that or I'm doing something wrong), so this shouldn't be merged until that's fixed.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants