tokihisu-cat-translate-0.8b-q4

CAT-Translate-0.8B Q4 ONNX for TOKIHISU

This repository contains a Transformers.js-compatible, symmetric Q4 ONNX export of cyberagent/CAT-Translate-0.8b for Japanese ↔ English translation.

• Source revision: b386247b30fd3b1911f7c694700d73843c376c6e
• Runtime task: text-generation
• Export task: text-generation-with-past
• Quantization: 4-bit symmetric MatMul weight-only quantization
• ONNX graph: 1.09 MiB
• External tensor data: 894.77 MiB

Transformers.js

import { pipeline } from '@huggingface/transformers';

const translator = await pipeline('text-generation', 'goryodog/tokihisu-cat-translate-0.8b-q4', {
  device: 'webgpu',
  dtype: 'q4',
});

const prompt = 'Translate the following Japanese text into English.\n\nこんにちは。';
const result = await translator([{ role: 'user', content: prompt }], {
  max_new_tokens: 160,
  do_sample: false,
  return_full_text: false,
});
console.log(result);

The source model requires the instruction format:

Translate the following {source language} text into {target language}.

{source text}

See conversion-manifest.json for exact package versions, source pin, graph I/O, quantization settings, and SHA-256 hashes.

License

The source model is MIT licensed. See LICENSE and the original model repository.