kaojai/quickmt-en-th-onnx

QuickMT English–Thai ONNX Q8

This repository contains a Q8 ONNX Runtime conversion of quickmt/quickmt-en-th, prepared by KaoJai.ai for browser-local translation experiments.

The package contains separate English source and Thai target tokenizers, plus quantized encoder and merged-decoder ONNX graphs. It was converted from the upstream EOLE model revision a4f1466054c237018111d97074b07e8006e6e2ca.

Intended runtime

Use ONNX Runtime Web with an autoregressive decoder loop. This package is not yet compatible with Transformers.js pipeline("translation"), because its merged decoder does not yet expose the key/value-cache interface that generic Transformers.js generation expects.

Files required by the browser runtime:

• onnx/encoder_model_quantized.onnx
• onnx/decoder_model_merged_quantized.onnx
• tokenizer.json and tokenizer_config.json for English source text
• target-tokenizer/tokenizer.json and target-tokenizer/tokenizer_config.json for Thai decoded text

License and attribution

The upstream QuickMT English–Thai model is published by quickmt under CC BY 4.0. KaoJai.ai modified the upstream model by converting its EOLE weights to Q8 ONNX Runtime graphs for browser-local inference. Attribution and source details are in ATTRIBUTION.md.

This conversion does not imply endorsement by quickmt.