Skip to content

release: tha-g2p-base-1.0 (gated ByT5-base student) - #13

Merged
ronaldtse merged 1 commit into
mainfrom
release/tha-g2p-base-1.0
Aug 19, 2026
Merged

release: tha-g2p-base-1.0 (gated ByT5-base student)#13
ronaldtse merged 1 commit into
mainfrom
release/tha-g2p-base-1.0

Conversation

@ronaldtse

Copy link
Copy Markdown

Summary

  • First distilled model release: tha-g2p-base-1.0 — ByT5-base student from the B-K/umt5 teacher (4.43% PER), sequence-level KD over 48,757 beam-4 labels
  • Gate: student 9.19% PER, +4.76pp ≤ +5pp budget (same harness as teacher: 1,219 Kaikki sentences, beam-4, corpus-level PER)
  • Distill runner hardening: resumable beam-4 labeling (OOM split-retry + token-budget batching), byte-length truncation in the training collate, gradient checkpointing, degenerate-label filtering, checkpoint+optimizer resume

Test plan

  • gate eval pass (evaluate_per, both models on same harness)
  • CI green
  • export fp32 -> parity gate -> GH Release + models.yaml entry (publish_model.py) -> production fetch verify

Sequence-level KD from B-K/umt5 teacher (4.43% PER): 48,757 beam-4
labels, gate student 9.19% (+4.76 <= +5pp). Distill runner hardened
along the way: resumable beam-4 labeling with OOM split-retry,
token-budget batching, byte-length truncation in the training collate
(the OOM saga's root cause), gradient checkpointing, degenerate-label
filtering, teacher recovery script with secryst-recipe stages.
@ronaldtse
ronaldtse merged commit 70065b2 into main Aug 19, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant