feat(distill): WO07 logit-distillation runner (Hebrew base->small case study) - #5
Merged
Conversation
added 2 commits
August 16, 2026 17:31
…yT5-small) Loss = 0.5 * T^2 * KL(teacher||student) + 0.5 * CE on the same hebrew-v4 corpus the teacher trained on; student initializes from google/byt5-small pretraining (pretrained backbone is essential — from-scratch ByT5 plateaus per the Thai ablations). Periodic step-checkpoints + auto-resume, A10G only, watchdog invocation in the module docstring. Teacher frozen in fp16; DER before/after to be reported in rababa RESULTS.md when the run completes.
A branch switch during an image build aborts it ('.git/HEAD was
modified during build process'); model zips have no business in the
image either.
ronaldtse
force-pushed
the
feat/imf-distill
branch
from
August 16, 2026 09:31
93a8518 to
72e8976
Compare
added 4 commits
August 16, 2026 20:02
One harness, same nakdimon test pairs, both models in fp16 on one GPU: the DER comparison for RESULTS.md is apples-to-apples, not against a number measured with a different evaluator/segmentation.
The kwarg crash-looped every watchdog launch (parity gates and the distill eval). Plain copy=True again; avoid branch switches while a build is in flight instead.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Work order TODO.runtime-arch/07: the logit-distillation runner that
shrinks models for client-side use — the edge tier of the size ladder
(server fp32 zips → edge distilled ByT5-small students int8
~110-250MB → browser tiny students ~60MB, the latter a config away).
First case study (per the work order): Hebrew ByT5-base s43 teacher →
ByT5-small student on the same
hebrew-v4corpus the teacher trainedon. Teacher frozen in fp16; loss =
0.5 * T^2 * KL(teacher || student) + 0.5 * CE(T=2). Studentinitializes from
google/byt5-smallpretraining — a pretrained backboneis essential (from-scratch ByT5 plateaus ~13% PER in the Thai ablations;
mode collapse resists all standard fixes).
Ops discipline (per the campaign hard rules):
newest (
rababa-checkpoints:/rababa_hebrew_distill_small/run-001)until modal run --detach src/gpu/modal_distill.py::main --spec heb-diac-small; do sleep 60; doneThe run is live now; DER before/after lands in rababa RESULTS.md when
it completes, then the student goes through the same WO02 export →
WO03 gate → release path as every other model.
Test plan