From 772c92f5123a9089b7f0ee2a1c73ae625cd1dd15 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Wed, 2 Sep 2026 13:32:12 +0200 Subject: [PATCH] =?UTF-8?q?feat(distill):=20G2b=20label-scale=20rung=20?= =?UTF-8?q?=E2=80=94=2048k=20units=20from=20full=20Tashkeela?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The domain-residual causal test: same G2a recipe, 5x the corpus (144,662 units from cleaned Tashkeela, classical + MSA). Gate >= 0.5pp over G2a's 4.5701. --- src/gpu/distill_specs.yaml | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/src/gpu/distill_specs.yaml b/src/gpu/distill_specs.yaml index 68c461a..ea0e65e 100644 --- a/src/gpu/distill_specs.yaml +++ b/src/gpu/distill_specs.yaml @@ -324,6 +324,27 @@ ara-diac-tiny-max: optimizer: muon note: gapless 30M rung - full data + r7 + Muon + 6ep; the title claim's decisive test +ara-diac-small-2-6ep-tashkeela: + # G2b: the domain-residual causal test (TODO.publish-client 01). + # Identical to the G2a recipe (r7 teacher, Muon, 6 epochs, vanilla + # ByT5-small) except the corpus: 144,662 units from the FULL cleaned + # Tashkeela corpus (~5x the r5-units corpus) — classical Arabic books + # + MSA. Gate: >= 0.5pp full-set DER improvement over G2a (4.5701) + # attributes the residual to domain coverage; flat reframes it. + teacher: rababa_arabic_byt5/run-007-news/best + teacher_volume: rababa + out_volume: rababa + student_init: google/byt5-small + train: tashkeela-units/train.txt + train_extra: [] + unit_limits: + - 48000 + max_len: 1450 + label_beams: '1' + out: rababa_arabic_distill_small/run-011-tashkeela-6ep + mode: sequence + optimizer: muon + note: G2b label-scale rung; 5x corpus, same levers; gate vs 4.5701 ara-diac-tiny-stitched: teacher: rababa_arabic_byt5/run-006-morph/best teacher_volume: rababa