Hugging Face's logo

Interscript
/
heb-diac-plane-1.0

ONNX
Hebrew
secryst
transliteration
diacritization
byt5
Model card Files
xet
 Community
heb-diac-plane-1.0
File size: 2,934 Bytes
e356123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
---
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.