docs: Hebrew layerdrop verdict — depth-cut collapse - #179
Merged
Conversation
Single-variable replication of the Arabic depth-cut on the Hebrew logit-KD lineage collapsed (+53.77pp vs full-depth's 30.38) despite normal training convergence. The compressible-depth finding is scoped: recipe- and language-dependent, not a universal property. Confound noted: Arabic used sequence-KD + Muon at 6ep; Hebrew used logit-KD at 3ep.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Result
heb-diac-small-s46-layerdrop full Nakdimon gate: teacher 23.72
(reproduces), student 77.48 DER (+53.77pp [51.64, 55.92], p=0) vs
the full-depth rung's 30.38. Training converged (val_loss 0.550) —
the collapse is at generalization.
Interpretation
The Arabic finding (depth-cut costs 1.21pp) does not transfer: a ~47pp
catastrophe here. Honest confound: recipe differs (Arabic =
sequence-KD + Muon + 6ep; Hebrew = logit-KD + 3ep), so the counter-
example may be depth-cut-under-logit-KD rather than language — but the
cross-lingual universal claim is dead either way. Paper B's depth
paragraph gets this as its scoping counterexample.