From b73892dbdec0e189110fe391fd92b8a8fccf8d13 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Thu, 3 Sep 2026 18:49:39 +0200 Subject: [PATCH] docs: ara-diac-small-2.1 README + registry pointer The 2.1 rung's release notes (4.5701 [1.91,2.35], epochs lever, frontier context) and the export registry entry now points at its own README instead of borrowing 2.0's. --- .../ara-diac-small-2.1.README.md | 29 +++++++++++++++++++ src/gpu/modal_export.py | 2 +- 2 files changed, 30 insertions(+), 1 deletion(-) create mode 100644 models/ara-diac-small/ara-diac-small-2.1.README.md diff --git a/models/ara-diac-small/ara-diac-small-2.1.README.md b/models/ara-diac-small/ara-diac-small-2.1.README.md new file mode 100644 index 0000000..6f42390 --- /dev/null +++ b/models/ara-diac-small/ara-diac-small-2.1.README.md @@ -0,0 +1,29 @@ +# ara-diac-small-2.1 + +Arabic diacritization (haraqat restoration). Client-tier ByT5-small +student — the 2.0 recipe (r7 canonical teacher, Muon) held for +**double the epochs** (6 instead of 3, 39,018 steps): the epochs lever +moves the rung 4.82 → **4.5701** full-set windowed DER-CE. + +The paired-bootstrap delta vs the teacher (same harness, seed 42, +1,000 resamples) is [1.91, 2.35] — disjoint from the 2.0 rung's +[2.36, 2.82]: the improvement is statistically real, and the frontier +separations hold end to end (1.0 8.26 → lite 5.78 → 2.0 4.82 → 2.1 +4.57 → teacher 2.29). + +The epochs result reframes the residual: doubling training buys +−0.25pp, so undertraining is not the dominant term — the E2/E3 domain +attribution substantially stands. The strict teacher+0.5pp gate +remains missed (disclosed); the causal follow-ups (E6 register swap — +negative; G2b 5× classical corpus — the add-direction test) are in +docs/RESULTS.md. + +```python +from interscript_ml import Model +model = Model.load("ara-diac-small-2.1") +model.translate("قوله فحكمها في الوفاة") +``` + +Same IMF v1 contract as 2.0: dynamic fetch, sha256-verified, KV +decode, margin report alongside the parity block. Checkpoint: +`rababa_arabic_distill_small/run-007-r7-muon-6ep/best`. diff --git a/src/gpu/modal_export.py b/src/gpu/modal_export.py index f3e6a1d..ab19052 100644 --- a/src/gpu/modal_export.py +++ b/src/gpu/modal_export.py @@ -127,7 +127,7 @@ "volume": "/volumes/rababa-checkpoints", "checkpoint": "rababa_arabic_distill_small/run-007-r7-muon-6ep/best", "metadata": "models/ara-diac-small/ara-diac-small-2.1.metadata.yaml", - "readme": "models/ara-diac-small/ara-diac-small-2.0.README.md", + "readme": "models/ara-diac-small/ara-diac-small-2.1.README.md", "test_volume": "/datasets/rababa", "test_data": "arabic-sadeed-imf/test.jsonl", "probe": "قوله",