opus-mt-tc-bible-big-poz-en โ€” CTranslate2 int8

An int8 CTranslate2 conversion of Helsinki-NLP/opus-mt-tc-bible-big-poz-en (Malayo-Polynesian โ†’ English, transformer-big, OPUS-MT / Tatoeba-Challenge release 2024-08-17; covers Indonesian, Malay and related languages).

Converted with:

ct2-transformers-converter (TransformersConverter, quantization="int8",
                            copy_files=["source.spm", "target.spm"])

Quality (measured on this int8 artifact)

FLORES-200 devtest, idโ†’en, 300 sentences, CPU: chrF2 69.3 at beam 1 (69.9 at beam 4). The original fp32 model reports chrF 0.684 on flores200-devtest, so int8 quantization is essentially loss-free.

Usage

import ctranslate2, sentencepiece as spm

translator = ctranslate2.Translator("model_dir", compute_type="int8")
sp_src = spm.SentencePieceProcessor("model_dir/source.spm")
sp_tgt = spm.SentencePieceProcessor("model_dir/target.spm")

tokens = sp_src.encode("Kehidupan seorang bankir itu sulit.", out_type=str) + ["</s>"]
result = translator.translate_batch([tokens], beam_size=1)
print(sp_tgt.decode(result[0].hypotheses[0]))

Attribution & license

Original model by the Language Technology Research Group at the University of Helsinki (OPUS-MT project), released under Apache-2.0. This repository redistributes a quantized conversion of those weights under the same license. Please cite the OPUS-MT and Tatoeba Translation Challenge papers.

Downloads last month
20
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for luigi000/opus-mt-tc-bible-big-poz-en-ct2-int8

Finetuned
(3)
this model