opus-mt-en-ml (int8 ONNX)

English โ†’ Malayalam machine translation. This is an int8-quantised ONNX export of Helsinki-NLP/opus-mt-en-ml (commit c31fa33320302c9cacd561792bb0d6b27d116cf4), made for on-device, offline translation in the Daily Whisper Android notes app.

The model weights are the work of the Language Technology Research Group at the University of Helsinki (OPUS-MT, Jรถrg Tiedemann et al.), released under the Apache License 2.0. This repo only changes the file format; it is redistributed under the same licence. No fine-tuning or training was done.

Original model

  • model: transformer-align (MarianMT), dataset: opus+bt+bt (2020-04-28)
  • pre-processing: normalization + SentencePiece
  • upstream benchmark (Tatoeba.en.ml): BLEU 19.1. Int8 quantisation may lower quality slightly; not re-benchmarked.

Files

The layout matches the Xenova/opus-mt-* repos, so it works with transformers.js-style loaders.

file bytes sha256
onnx/encoder_model_quantized.onnx 32,271,107 43276f6a5e69efa0377b82adc9dff0ae018be340c4f4a179b6cd74e0d66f5365
onnx/decoder_model_merged_quantized.onnx 51,980,871 ed69a7e5e7461fb3029c1aadc254fedaee9abdc1c5a462002b51dcbce2740be1
tokenizer.json 1,022,410 f129181a0e41b62c67b44e5d0bcb4585e9a6edfa5748f2043e5fada2e6e27068
config.json 1,391 5fc9dab47f3453afa6d064e13433b4e836a3a71cbf4da84db8ec29486cd08949

How it was made

  1. Exported to ONNX with Hugging Face Optimum 1.23.3 (task text2text-generation-with-past, transformers 4.46.3, torch 2.5.1), producing encoder_model.onnx and decoder_model_merged.onnx (decoder with KV cache, merged with an use_cache_branch switch).
  2. Quantised with ONNX Runtime 1.28.0 quantize_dynamic (weight type QInt8, per-channel, reduce_range).
  3. tokenizer.json generated from the Marian SentencePiece vocab with the transformers.js Marian converter; config.json copied from upstream (generation defaults max_length, num_beams, bad_words_ids kept in it, as in the Xenova repos).

Known limitations

  • Same as the upstream model: trained on OPUS data, short sentences work best, quality varies by domain.
  • Malayalam text mixed with English words (code-mixed / "Manglish") translates poorly.
  • Not suitable for high-stakes translation.

Citation

@InProceedings{TiedemannThottingal:EAMT2020,
  author = {J{\"o}rg Tiedemann and Santhosh Thottingal},
  title = {{OPUS-MT} โ€” {B}uilding open translation services for the {W}orld},
  booktitle = {Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT)},
  year = {2020},
  address = {Lisbon, Portugal}
}
Downloads last month
151
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for akm9191/opus-mt-en-ml-int8-onnx

Quantized
(2)
this model