From db28e87e57ce0922c4fc79ec92153fb2132fb6cd Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Fri, 28 Aug 2026 19:15:27 +0800 Subject: [PATCH] fix: remove duplicate ara-diac-tiny spec (last-key-wins hazard) PR #63 added a second ara-diac-tiny key; the pre-existing spec later in the dict silently won. Keep the existing clean-label spec (snapshot labels, gate <= 3.07). --- src/gpu/modal_distill.py | 34 ---------------------------------- 1 file changed, 34 deletions(-) diff --git a/src/gpu/modal_distill.py b/src/gpu/modal_distill.py index 4c38a9d..1094499 100644 --- a/src/gpu/modal_distill.py +++ b/src/gpu/modal_distill.py @@ -141,40 +141,6 @@ def _ensure_src_path() -> None: "mode": "sequence", "note": "r6 canonical (2.5793 DER); gate <= 3.07 windowed DER-CE", }, - "ara-diac-tiny": { - # the browser/Worker tier: ~45M params -> ~45MB int8 zip. Same - # teacher (r6), corpus, and frozen labels as run-002 — the ONLY - # variable is student capacity. From-scratch risk is real (Thai - # ablation: scratch byt5-small plateaued ~13% PER) but this is - # dense CE on 30k paragraph units, not RL; gate: does not - # collapse (finite output, DER bounded well below scratch) and - # <= 5.0 windowed DER-CE would make it shippable. - "teacher": "rababa_arabic_byt5/run-006-morph/best", - "teacher_volume": "rababa", - "student_config": { - "vocab_size": 384, - "d_model": 384, - "d_ff": 1024, - "d_kv": 64, - "num_layers": 6, - "num_decoder_layers": 2, - "feed_forward_proj": "relu", - "decoder_start_token_id": 0, - "eos_token_id": 1, - "pad_token_id": 0, - }, - "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_tiny/run-001", - "labels_file": "teacher_labels_v2.jsonl", - "labels_complete": "true", - "mode": "sequence", - "note": "tiny-tier probe: r6 teacher, run-002 corpus/labels, " - "~45M student; collapse check + DER gate <= 5.0", - }, "ara-diac-small-pkm": { # TODO.qwen-next/02 — the LongCat/Qwen capacity axis: keep the # ByT5-small compute, add product-key lookup memory (+~25M