Repository navigation
docs: tune the consistent-scaling ASTRA row; name the CG row EKFAC + CG - #570
Merged
Merged
Conversation
The original EK-FAC + ASTRA run (LDS 0.643, QLD 0.074, per-token 1.082) sits beside the CG, paper and consistent scaling rows. Its config loads the published EK-FAC factors, summed over the 4,608 training chunks. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
luciaquirke
changed the title
…nd spell EK-FAC Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
luciaquirke
changed the title
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
luciaquirke
changed the title
Relative damping 0.1 and lr 3e-6, the step size with the lowest mean final / |step-0| objective over test[50:60] from a 1e-3 to 1e-7 grid. EK-FAC + ASTRA (consistent scaling): LDS 0.579 [0.554, 0.601], QLD 0.073 [0.061, 0.087]. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The paper replication version of ASTRA uses scaling that was almost certainly unintended but that happens to make performance good on the leaderboard (this isn't super surprising - when we used the ~randomly selected CG solver instead that made performance even better).
Here we include both a row using the replication settings, and another row that recreates what I guess they intended ASTRA to be - consistent scaling, and a learning rate tuned using the solver error (not LDS or anything that requires retrains).
For our "sensible" row we use a robust damping heuristic (damping value = 0.1 * mean eigenvalue) used in other Grosse-supervised papers rather than the absolute damping value used in the ASTRA paper. The tuned lr is 3e-6. We use held-out tuning queries for this lr, although in practice you would probably want to use your actual intended queries. Here it doesn't make a difference to the optimal lr.