From 17c0a3b6fd324e67560436308b25730bec6a2cb6 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Tue, 1 Sep 2026 17:48:24 +0200 Subject: [PATCH] =?UTF-8?q?feat(distill):=20ara-diac-tiny-max=20=E2=80=94?= =?UTF-8?q?=20the=20gapless=2030M=20title=20test?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Full corpus (24k+6k vs the collapse rungs' 12k), r7 labels, Muon, 6 epochs at d384 scratch. The title claim's decisive experiment: conclusion-from-evidence, whichever way it lands. --- src/gpu/distill_specs.yaml | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+) diff --git a/src/gpu/distill_specs.yaml b/src/gpu/distill_specs.yaml index 67e0583..68c461a 100644 --- a/src/gpu/distill_specs.yaml +++ b/src/gpu/distill_specs.yaml @@ -293,6 +293,37 @@ ara-diac-small-layerdrop-6ep: mode: sequence optimizer: muon note: A2 rung; gate tracks G2a (4.57) within 0.1pp at half depth +ara-diac-tiny-max: + # THE gapless title test: the 30M class with the campaign's FULL + # lever set — full corpus (24k+6k, not the 12k subset the collapse + # rungs used), r7 labels, Muon, 6 epochs. If this still collapses, + # the "30M encoder" claim dies by our own standard; if it works, + # data quality rescued the small model and the title stands. + teacher: rababa_arabic_byt5/run-007-news/best + teacher_volume: rababa + out_volume: rababa + student_config: + d_model: 384 + d_kv: 64 + d_ff: 1536 + num_heads: 6 + enc_layers: 8 + dec_layers: 8 + train: r5-units/domain.txt + train_extra: + - r5-units/replay.txt + unit_limits: + - 24000 + - 6000 + max_len: 1450 + label_beams: '1' + out: rababa_arabic_distill_small/run-010-tiny-max + labels_file: teacher_labels_r7.jsonl + labels_complete: 'true' + mode: sequence + optimizer: muon + note: gapless 30M rung - full data + r7 + Muon + 6ep; the title + claim's decisive test ara-diac-tiny-stitched: teacher: rababa_arabic_byt5/run-006-morph/best teacher_volume: rababa