Skip to content

feat(distill): tiny client tier — byte-vocab students + microkimi bridge KD - #23

Merged
ronaldtse merged 6 commits into
mainfrom
feat/tiny-tier-reconstruction
Aug 22, 2026
Merged

feat(distill): tiny client tier — byte-vocab students + microkimi bridge KD#23
ronaldtse merged 6 commits into
mainfrom
feat/tiny-tier-reconstruction

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Summary

  • student_config: custom byte-vocab T5 students (33M/70M rungs) — byte vocab makes embeddings negligible, so small transformers are almost all compute (~30MB int8 artifacts)
  • distill_microkimi: calibration → ridge-solved frozen projectors → CE + hidden-MSE training; measured effect on the from-scratch collapse: 75.8 → 71.12 PER (Thai, 33M)
  • out_volume: student outputs on the secryst volume — the rababa volume serves torn reads to subsequent containers
  • labels_complete/labels_file/torn-view guard; hardened Step-2 label loading
  • Arabic switched to the r6 canonical teacher (2.5793 DER, per the 2026-08-21 manifest)
  • Persian datasets volume + data_volume override (six-mount shadowing workaround); jsonl corpus limits

Test plan

  • syntax + ruff
  • 70M bridge rung gate (in flight), then export int8 → parity → publish

Ronald Tse added 6 commits August 21, 2026 23:58
…imi bridge KD
Reconstruction after an accidental reset --hard wiped the uncommitted
work (running containers kept their snapshots; this restores source).
- student_config: custom T5Config byte-vocab students (33M/70M rungs)
- distill_microkimi: calibration -> ridge-solved frozen projectors ->
CE + hidden-MSE (rescues the from-scratch collapse: 75.8 -> 71.12 PER
at 33M; 70M rung in flight)
- out_volume: outputs on the secryst volume (rababa volume serves torn
reads cross-container)
- labels_complete / labels_file / torn guard; hardened Step-2
- r6 canonical Arabic teacher; persian datasets volume + data_volume
override (six-mount shadowing workaround); jsonl corpus limits
Measured quantization deltas on khm: fp16 ~0.43pp, int8 ~0.84pp against
the flat 0.2pp bar, so quantized client zips could never pass strict.
Limits keyed on metadata.precision; unknown precision falls back to the
fp32 bar.
…ation
Re-measured run-003 on the beam-4 harness (12.06% PER, EM 87.94%,
+7.63pp over the 4.43% teacher, n=1219). Capacity-limited but the
smallest non-collapsing rung; exported at int8 (~300MB) per the
frontier table. Frontier row updated 12.63 -> 12.06 (full-label run).
ml-models -> interscript/interscript-ml everywhere (index URLs, metric
sources, workflow provenance subject, script defaults, Modal container
paths); interscript-ml-train app renamed; root tooling package renamed
interscript-ml-tools (runtime keeps the interscript-ml pip name); npm
manifest package marked deprecated, never published, license BSD-3;
TODO.distribution names mechanically updated + historical banner.
Old names survive only inside deprecation notes explaining them.
ronaldtse pushed a commit that referenced this pull request Aug 22, 2026
metrics-sources pins this section at ref release/tha-g2p-small-1.0
until it lands on main via PR #23 (fas-g2p release-branch pattern).
@ronaldtse
ronaldtse merged commit c80a78e into mainAug 22, 2026
16 of 17 checks passed
Sign up for freeto 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

@ronaldtse