KazakhBERTmulti-onnx

This repository provides a ready-to-use ONNX model directory for @sarmay/kaz-converter-lm.

It is intended to improve ambiguous Arabic Tote Zhazu -> Cyrillic conversion in Node.js by scoring candidate Kazakh sentences with a masked language model.

Model repo:

• https://huggingface.co/sarmay/KazakhBERTmulti-onnx

Compatible npm packages:

• @sarmay/kaz-converter
• @sarmay/kaz-converter-lm

Use with npm

Install:

npm install @sarmay/kaz-converter @sarmay/kaz-converter-lm

Download this model repository to a local directory such as:

models/KazakhBERTmulti-onnx

If you only want the quickest path, use the bundled CLI:

npx sarmay-kaz-download ./models/KazakhBERTmulti-onnx

Then use:

import { createOnnxArabicToCyrillicConverter } from "@sarmay/kaz-converter-lm";

const converter = await createOnnxArabicToCyrillicConverter({
  modelDirectory: "./models/KazakhBERTmulti-onnx"
});

console.log(await converter.convertAsync("الما بار."));
console.log(await converter.convertAsync("بىر كۇنى"));
console.log(await converter.convertAsync("اكەم كەلدى."));

Expected local layout

models/
  KazakhBERTmulti-onnx/
    model.onnx
    tokenizer.json
    tokenizer_config.json
    special_tokens_map.json
    config.json
    vocab.txt

Expected files

This repository should contain:

• model.onnx
• tokenizer.json
• tokenizer_config.json
• special_tokens_map.json
• config.json
• vocab.txt

Minimum required for loading:

• model.onnx
• tokenizer.json
• tokenizer_config.json

Source model

This ONNX export is derived from:

• amandyk/KazakhBERTmulti

and exported for local Node.js inference with:

• onnxruntime-node
• @huggingface/tokenizers

Intended use

Use this model when you want:

• a local Node.js scorer
• better disambiguation than pure rule conversion
• a drop-in model directory that works with createOnnxArabicToCyrillicConverter()

This repository does not package the model into npm. Download the model files separately and pass the directory path to the LM package.

Intended package pair

• Core package: @sarmay/kaz-converter
• LM package: @sarmay/kaz-converter-lm