File size: 2,664 Bytes
61cf91f | 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 | ---
license: bsd-3-clause
language: ar
inference: false
tags:
- secryst
- transliteration
- diacritization
- onnx
- byt5
---
# ara-diac-2.0
Part of the [Secryst](https://secryst.github.io) phonological-layer family — a dedicated model for **diacritization** (AR), distributed as a Secryst IMF artifact and loadable through any Secryst crystal:
```python
from secryst import Model
Model.load("ara-diac-2.0")
```
- Artifact: `ara-diac-2.0-fp32.zip` (2.59 GiB)
- Precision variants: `ara-diac-2.0-int8`
## 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 | 2.2864 | rababa/docs/RESULTS.md#r7-verdict-table-sadeeddiac-25-2026-08-28 |
| der_morph_greedy | 1.3343 | rababa/docs/RESULTS.md#r7-verdict-table-sadeeddiac-25-2026-08-28 |
| wer_wikinews_multiref | 17.3794 | rababa/docs/RESULTS.md#r7-ood-verdict-table-wikinews-2024-multiref-2026-08-28 |
| der_wikinews_multiref | 11.8273 | rababa/docs/RESULTS.md#r7-ood-verdict-table-wikinews-2024-multiref-2026-08-28 |
## Benchmark
SadeedDiac-25 — full 1,200-paragraph set, Misraj's evaluator, zero
skipped examples, greedy decode (protocol-matched published numbers
only):
| system | Total DER | kind |
|---|---|---|
| Claude-3.7-Sonnet | 1.39% | frontier LLM |
| **ara-diac-2.0 (this model)** | **2.29%** | dedicated, 580M |
| GLM-5.2 | 2.69% | frontier LLM |
| Gemini-Flash-2.0 | 3.19% | frontier LLM |
| GPT-4 | 3.86% | frontier LLM |
| Sadeed-1.5B | 7.29% | leaderboard baseline |
Word-final iʿrāb endings (the hardest component): 1.33% DER.
Out-of-domain generalization (WikiNews-2024 multi-reference):
17.38 WER / 11.83 DER.
## Architecture
ByT5-large-class seq2seq transformer (580M parameters), byte-level
input/output (no tokenizer — id = byte + 3). Trained on classical and
modern Arabic running text with a news-domain adaptation stage over a
morphology-auxiliary curriculum. The canonical distribution is the
Secryst IMF artifact (sha256-verified, parity-gated against the
PyTorch reference); this repository mirrors those exact bytes.
## 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-ml) 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.
|