Repository navigation
Add projected activation covariance preconditioner leaderboard row - #569
Merged
Merged
Conversation
The validate step reads query.aggregation, which DataConfig lacks. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
activation_baseline.py takes --hessian, which whitens each module's pooled activations by its EK-FAC activation factor, and --projection_dim, which applies bergson's per-module random projection. Adds the projection-64 leaderboard row, the reproduction steps for both preconditioned rows, and their filter configs. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
luciaquirke
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds the projected and preconditioned activation similarity row to the leaderboard.
The preconditioner is the dampened activation covariance matrix:
(A + damping * mean(eig A) I)^-1/2.Ais the activation covariance matrix (the module's EK-FAC activation factor). The projection is the right-side of bergson's per-module (TrackStar-style) random projection.Also builds the baseline LDS query as a
QuerySetConfig: the validate step readsquery.aggregation, whichDataConfiglacks.