heb-diac-plane-2.0

Hebrew nikud diacritization, plane-factorized (run-022): ByT5-base encoder + per-character nikud-class head, K=3 mask-predict decode, int8 (dynamic QInt8) ONNX, 416 MB. Trained on hebrew-v4 (nakdimon + sefaria + distilled + expanded) on HF Jobs.

Metrics (runtime protocol)

benchmark	value
nakdimon test DER (greedy, int8, py runtime)	8.18 (1.0: 12.48)
phonikud heb-g2p-benchmark CER / no-stress (via nikud_ipa rules)	0.2393 / 0.1389

Features

• translate(text, preserve_diacritics=True) — user-supplied nikud is pinned inside the decode and round-trips byte-exactly (leading marks included).
• Language-generic plane contract (kind: plane); same runtime as the Arabic plane family.
• nikud→IPA composition: interscript.ml.nikud_ipa.plane_to_ipa.

Provenance

• Training: TODO.sota/04 (run-022, HF Jobs a100-large); corpus Interscript/hebrew-v4 (private); recipe mirrors run-021 at byt5-base/4ep/K=3.
• Artifact sha256: 9c8f0432294ebae6abc847ddb79572c1ab979f4851f03d44a9bb6495ba767a9e.
• License: BSD-3-Clause (code + weights).

Mirror provenance

Canonical channel: GitHub Releases (interscript/interscript-models), sha256 9c8f0432294ebae6abc847ddb79572c1ab979f4851f03d44a9bb6495ba767a9e. This HF repo is a 1:1 mirror of the index entry heb-diac-plane-2.0.