ara-diac-2.0
Part of the Secryst phonological-layer family โ a dedicated model for diacritization (AR), distributed as a Secryst IMF artifact and loadable through any Secryst crystal:
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 โ 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 iสฟrฤb 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 โ 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 and published through the 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.