feat(distill): SVD width-stitch + ara-diac-tiny-stitched rung - #73
Merged
Conversation
svd_stitch_state projects pretrained wide weights into a narrow state dict preserving the top singular subspace per tensor; the config builders stitch instead of random-initing when a spec sets both student_config and student_init. New rung: same d384 width as the collapsed from-scratch run-005, init from pretrained ByT5-small (~27M params = ~27MB int8), clean r6 labels, gate <= 3.07. Also records the capacity-law evidence arc in the distribution history.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The capacity law (from-scratch collapses; init is the variable) now
has clean Arabic evidence (run-005: 74.68 vs gate 3.07). This rung
changes exactly that variable: same d384 width, same clean labels,
same gate - but SVD-stitched down from pretrained ByT5-small
(d1472 enc12/dec4 gated-gelu; ~27M params = ~27MB int8, the browser
budget).
(2-D shrink each dim independently; 1-D/N-D slice), unit-tested
for shape coverage and subspace-action preservation
feed_forward_proj becomes a passthrough (gated-gelu needed to
match ByT5 FF weight names)
TODO.distribution/00-overview.md for paper/history