tashkeel.onnx - Arabic diacritisation (tashkeel) model, 4,788,213 bytes

Model: the bundled ONNX model of libtashkeel by Musharraf Omer,
       https://github.com/mush42/libtashkeel (file crates/core/data/ort/model.onnx,
       commit 9b91c2dbc3dc4085dc4f77a9e250e5dd095d7d0c), taken unmodified and renamed.
       libtashkeel, including the bundled model, is licensed under the MIT License
       (LICENSE-libtashkeel-MIT.txt).
Training code: Hareef ("Sarf" model: GRU + transformer encoder, trained mainly on
       Modern Standard Arabic) by Musharraf Omer, https://github.com/mush42/hareef
       (commit fdb579a7cfe987430d172c199b4d038760f6a8e7), MIT License
       (LICENSE-hareef-MIT.txt).
SHA-256: 52c40d42ca094ff9ff07a091d0a1eafd7baa5cf2388e60a9cc89bacd8e5e955f

Pre- and post-processing (character and diacritic vocabularies, hint encoding,
annotation) are re-implemented in TypeScript in src/processors/text/tashkeel.ts
from libtashkeel's crates/core/src/lib.rs; the vocabularies are copied unmodified
from crates/core/data/*.json into this folder (input_id_map.json, target_id_map.json,
hint_id_map.json) and tests/tashkeel.test.ts checks the TypeScript tables against them.

The file is served unmodified from /tashkeel/tashkeel.onnx by
scripts/copy-tashkeel.mjs. Inference uses the ONNX Runtime Web (MIT, Microsoft)
WebAssembly runtime already copied to /bg-removal/ by scripts/copy-bg-removal.mjs.