--- license: cc-by-4.0 language: - fr - ar pipeline_tag: translation library_name: transformers.js base_model: Helsinki-NLP/opus-mt-fr-ar tags: - translation - marian - opus-mt - onnx --- # OptimaMT fr → ar (French to Arabic) Drop-in copy of an ONNX export of [Helsinki-NLP/opus-mt-fr-ar](https://huggingface.co/Helsinki-NLP/opus-mt-fr-ar) at revision `5469dc759cd9` (same layout as the Xenova exports) with a lighter merged decoder: ONNX Runtime's fast mode (`disable_quant_qdq = 1`, 2.3–2.5× faster) at half the memory. Same weights, same tokens. Part of [OptimaMT](https://github.com/AmerAlreyahi/OptimaMT): [all 51 pairs](https://huggingface.co/collections/OptimaMT/optimamt-lighter-opus-mt-onnx-models-6af228dc9cd91f015e56e869), named `OptimaMT/-`. **Used in production:** - [OpenCircle](https://github.com/AmerAlreyahi/OpenCircle), a desktop screen translator for Windows and Linux, ships these models since v0.1.4. | Memory after load | Linux | Windows | Identical tokens | |---|---:|---:|---| | `disable_quant_qdq = 1` | 525 → 281 MiB (-47 %) | 500 → 257 MiB (-49 %) | yes | | ONNX Runtime defaults | 281 → 220 MiB (-22 %) | 196 → 165 MiB (-16 %) | yes | FLORES-200 devtest (1,012 sentences): the same translations as the original, BLEU 13.61, chrF 45.82 (greedy). ONNX Runtime 1.31.1, CPU, 4 threads (Linux; Windows in a VM). Interactive results: [OptimaMT/results](https://huggingface.co/spaces/OptimaMT/results) · method: [REPORT.md](https://github.com/AmerAlreyahi/OptimaMT/blob/main/REPORT.md). ## Use ```js import { pipeline } from "@huggingface/transformers"; const translator = await pipeline("translation", "OptimaMT/fr-ar", { dtype: "q8", session_options: { extra: { session: { disable_quant_qdq: "1" } } }, }); const [out] = await translator(">>ara<< Le phare se dresse sur la côte."); console.log(out.translation_text); ``` Python ([translate.py](https://github.com/AmerAlreyahi/OptimaMT/blob/main/scripts/translate.py): onnxruntime, sentencepiece, numpy): ```bash python translate.py OptimaMT/fr-ar "Le phare se dresse sur la côte." --lang ">>ara<<" ``` Start every input with `>>ara<<` (the target variant). Use transformers.js 3.x: 4.3.2 splits this token, with the original export too. ## Files `onnx/decoder_model_merged_quantized.onnx` is rebuilt; `tokenizer.json`, `special_tokens_map.json` and `tokenizer_config.json` were written in the Xenova layout; every other file is the original, byte for byte. Vocabulary: 62,419 tokens. ## Licence and credits - Weights: [OPUS-MT](https://huggingface.co/Helsinki-NLP/opus-mt-fr-ar) by the University of Helsinki, CC BY 4.0. - ONNX export: an ONNX export of [Helsinki-NLP/opus-mt-fr-ar](https://huggingface.co/Helsinki-NLP/opus-mt-fr-ar) at revision `5469dc759cd9` (same layout as the Xenova exports). - Modified: the decoder's output projection was moved out of its two branches (graph change only). CC BY 4.0.