| --- |
| license: bsd-3-clause |
| language: he |
| inference: false |
| tags: |
| - secryst |
| - transliteration |
| - diacritization |
| - onnx |
| - byt5 |
| --- |
| |
| # heb-diac-plane-1.0 |
|
|
| Part of the [Secryst](https://secryst.github.io) phonological-layer family — a dedicated model for **diacritization** (HE), distributed as a Secryst IMF artifact and loadable through any Secryst crystal: |
|
|
| ```python |
| from secryst import Model |
| Model.load("heb-diac-plane-1.0") |
| ``` |
|
|
| - Artifact: `heb-diac-plane-1.0.zip` (0.20 GiB) |
|
|
| ## Measured quality |
|
|
| Full evaluation sets, zero skipped examples, the benchmark's own evaluator. Provenance for every number is the results record of the training project. |
|
|
| | metric | value | provenance | |
| |---|---|---| |
| | der_total_greedy | 12.48 | interscript/interscript-models docs/RESULTS.md#run-021-heb-plane--1248-nakdimon-der-the-plane-family-generalizes-seq2seq-student-beaten-at-396pp-2026-10-07 | |
|
|
| ## Benchmark |
|
|
| Nakdimon test set, greedy decode, `seq2seq_der` — the exact protocol of |
| the shipped seq2seq student: |
|
|
| | model | Total DER | notes | |
| |---|---|---| |
| | **heb-diac-plane-1.0 (this model)** | **12.48%** | **219 MB int8** | |
| | heb-diac-1.1 (seq2seq student) | 16.44% | larger artifact | |
| | DictaBERT (fine-tuned BERT baseline) | ~35.6% | per-protocol reference | |
|
|
| −3.96pp (24% relative error reduction) over the shipped student at the |
| plane's on-device economics. Median per-example DER 9.8%; |
| 223,872 scored positions. |
|
|
| ## Architecture |
|
|
| Plane-factorized nikud restoration: a ByT5-small-class byte-level |
| encoder (no decoder) with a nikud-plane embedding and one |
| classification head per character position over 135 order-preserving |
| nikud classes. Decoding is K=2 mask-predict passes — fully parallel |
| forwards with the predicted plane fed back; no KV cache, no |
| autoregressive loop. Skeleton round-trip is byte-exact; nikud cluster |
| order is preserved as written (the corpus has no uniform canon). |
|
|
| The artifact is a single self-contained int8 ONNX graph in the plane |
| zip contract (metadata.yaml with member sha256s + plane.onnx + |
| classes.json), loadable by `interscript[ml]` `PlaneModel.from_zip`. |
| The canonical distribution is the sha256-verified release artifact; |
| this repository mirrors those exact bytes. |
|
|
| Provenance: interscript-train run-021 — v4 combined corpus (nakdimon + |
| sefaria + distilled v1/v2 + expanded_v2), 50,303 units, 2 epochs. The |
| verdict ledger records a disclosed harness correction (window-stitching |
| bug; first read 32.64%, corrected 12.48%). |
| |
| |
| ## License |
| |
| BSD-3-CLAUSE |
| |
| ## Provenance |
| |
| Trained in the [Secryst training monorepo](https://github.com/secryst/secryst-train) and published through the [interscript-ml](https://github.com/interscript/interscript-models) distribution contract (models.yaml index, sha256-verified release assets, parity-gated artifacts). The canonical download channel is GitHub Releases; this repository is the Hugging Face mirror of the same bytes, pinned to the index revision. |
| |
