File size: 1,482 Bytes
a7a7764 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 | ---
license: bsd-3-clause
tags:
- hebrew
- diacritization
- nikud
- plane-factorized
- onnx
- int8
library_name: onnx
---
# 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`.
|