tha-lexicon-kaikki-1.0

Thai word->IPA lexicon for the hybrid fast-path (WO03): 17,434 munch entries (>=2 chars) built from the Kaikki extraction of Wiktionary (kaikki.org/dictionary/Thai). License: CC BY-SA 3.0 — attribution: Wiktionary contributors / Kaikki project. Bench: benchmarks/thai_hybrid_bench.py (hybrid 0.1478 PER @ 0.73 ms/utt).

index-v9

Model index v9: adds heb-diac-plane-1.0 and heb-diac-plane-2.0 (ByT5-base, K=3, int8, nakdimon DER 8.18). 28 models.

heb-diac-plane-2.0

Hebrew plane scale-up (run-022, ByT5-base, K=3, int8, 416MB). nakdimon DER 8.18 (runtime protocol; 1.0 = 12.48). Card: cards/heb-diac-plane-2.0.md. sha256 sidecar alongside.

heb-diac-plane-1.0

heb-diac-plane-1.0 — nikud plane-factorized diacritization (on-device frontier)

First Hebrew plane model: the Arabic run-018 recipe applied unchanged
(ByT5-small-class byte-level encoder, no decoder; nikud-plane embedding;
135 order-preserving classes; K=2 mask-predict; parallel forwards, no
KV cache). One self-contained int8 ONNX graph per the plane zip
contract (metadata.yaml with member sha256s + plane.onnx +
classes.json), loadable by interscript[ml] PlaneModel.from_zip.

Nakdimon test, greedy, seq2seq_der (the exact heb-diac-1.1 protocol):

• Total DER 12.48% — the shipped seq2seq student scores 16.44 (−3.96pp,
  24% relative error reduction) at 219 MB-class on-device economics
• Median per-example DER 9.8%; 223,872 scored positions

Provenance: interscript-train run-021 (rababa_hebrew_plane), v4 combined
corpus (nakdimon + sefaria + distilled v1/v2 + expanded_v2; 50,303
units, 2 epochs), checkpoint classes=135. Verdict note: the first
reported number (32.64%) was a window-stitching harness bug; see the
interscript-models RESULTS ledger for the disclosed correction.
License: BSD-3-Clause.

index-v8

Interscript ML model index (index-v8).

Content-identical to index-v7 except canonical URLs: the repo was
renamed interscript-ml -> interscript-models (mirrors interscript-maps).
All entry URLs + provenance anchors point at the new name.

Runtimes: pin this index at your next release; older pins keep working
via the rename redirect.

golden-v2 — plane-family parity corpus

golden-v2 — neural parity corpus additions

Adds the plane-family golden sets (Python-reference outputs):

• ara-diac-plane-1.0.jsonl — 50 rows, SadeedDiac-25 skeletons
• ara-diac-plane-large-1.0.jsonl — same inputs, large model

The quantized-tolerance contract

Int8 graphs are not byte-portable across onnxruntime builds: near-tied
logits flip argmax. Measured py-reference vs ruby (onnxruntime gem
0.11.6, arm64): 3/50 rows differ with 1-2 char diffs on 600-1000-char
rows. Cross-runtime parity for quantized models is therefore:
per-row char agreement >= 99.5%, corpus agreement >= 99.9%.
Deterministic graphs (maps, fp32 fixtures) remain byte-exact.

index-v7

Interscript ML model index (index-v7).

Adds since index-v6:

• ara-diac-plane-1.0 — plane-factorized small (219 MB int8, full-set 3.5905 DER)
• ara-diac-plane-large-1.0 — plane-factorized large (866 MB int8, full-set 2.7397 DER, dedicated #2)

New field: kind: plane for plane-factorized artifacts (resolved by the
plane runtime; loader contract = metadata.yaml + plane.onnx +
classes.json with member sha256s).

Runtimes resolve DEFAULT_INDEX_URL against this release asset + sha256
sidecar.

ara-diac-plane-large-1.0

ara-diac-plane-large-1.0 — plane-factorized large encoder (dedicated #2)

The ByT5-LARGE plane variant: encoder-only plane factorization over 54
haraqat classes with K=2 mask-predict decoding, from rababa run-019 at
2 epochs (551k unique units).

Full-set SadeedDiac-25 (1,200 paragraphs, zero skipped, greedy):

• Total DER 2.7397 / Morph 1.6717 / Total WER 8.6083
• Dedicated podium: ara-diac-2.0 (teacher) 2.2864 > this model 2.7397 >
  plane-large 1-epoch 3.0078 > Gemini-Flash 3.1926
• Best dedicated system that still runs as one self-contained int8 ONNX
  graph (866 MB) with parallel K-pass decoding (no KV cache)

Provenance: run-019 epoch-2 final weights, exported on Modal from the
volume-native checkpoint (transit-tearing workaround); ONNX int8 sha256
f9cdba7448ddf32bfa3c50e5076dd9a8319d7b0ada01ecda6da2620e43abf428.
Process note: the epoch-2 number above is the corrected measurement
(the first E2 eval re-scored epoch-1 predictions; see docs/RESULTS.md).
License: BSD-3-Clause.

ara-diac-plane-1.0

ara-diac-plane-1.0 — plane-factorized Arabic diacritization (on-device frontier)

First release of the plane-factorized architecture: a ByT5-small-class
ENCODER with a diacritic-plane embedding and a per-position
classification head over 54 haraqat classes, decoded with K=2
mask-predict passes (host-side loop, no KV cache, fully parallel
forwards). The artifact is a single self-contained int8 ONNX graph per
the plane zip contract (metadata.yaml + member sha256s + plane.onnx +
classes.json), loadable by secryst-py PlaneModel.from_zip.

Full-set SadeedDiac-25 (1,200 paragraphs, zero skipped, greedy):

• Total DER 3.5905 / Morph 2.2149 — rung gate cleared (-0.98pp vs the
  seq2seq student's 4.5701)
• 219 MB int8 (-45% vs the shipped 491 MB seq2seq int8static)
• ~1,145 ms/window CPU vs ~3,459 ms for the seq2seq artifact (-67%)

Provenance: rababa run-018 (551k unique units, 2 epochs), checkpoint
sha256 26f49bd0ebbf3ca93cebf2f36bdd8b488f6bb17e99e3f61f9613e53d1852862.
License: BSD-3-Clause.

index-v6

Interscript ML model index (index-v6). Runtimes resolve DEFAULT_INDEX_URL against this release asset + sha256 sidecar.