-
Notifications
You must be signed in to change notification settings - Fork 561
[Proposal] Direct Logit Attribution Tool #1263
Copy link
Copy link
Open
Labels
complexity-moderateModerately complicated issues for people who have intermediate experience with the codeModerately complicated issues for people who have intermediate experience with the codeenhancementNew feature or requestNew feature or requestgood first issueGood for newcomersGood for newcomershelp wantedExtra attention is neededExtra attention is neededminorRelease a minor versionRelease a minor version
Metadata
Metadata
Assignees
Labels
complexity-moderateModerately complicated issues for people who have intermediate experience with the codeModerately complicated issues for people who have intermediate experience with the codeenhancementNew feature or requestNew feature or requestgood first issueGood for newcomersGood for newcomershelp wantedExtra attention is neededExtra attention is neededminorRelease a minor versionRelease a minor version
Proposal
PR #466 was created with the idea of adding a specific set of functions or utilities that could do the DLA process in a single call. The PR has been stale for two years now, but the idea behind it is sound.
Review the PR and create a new tool in
transformer_lens/tools/analysis/direct_logit_attribution.pythat handles this for the newTransformerBridgesystem.