Fairseq
lastmod 2020-10-28created 2020-10-281 edits
Fairseq(-py) is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling and other text generation tasks.
M2M-100
- Introducing the First AI Model That Translates 100 Languages Without Relying on English - About Facebook - Facebook open-sources a model that can translate between 100 languages without relying on English
- M2M-100 is an MMT (Multilingual Machine Translation) model, the first able to translate between 100 language pairs without relying on English data.
- Most translation goes through English, which has the most training data, but this model translates directly and so preserves meaning better
- Shows results more than 10% better on BLEU (Bilingual Evaluation Understudy) scores.
- Trained on a total of 2,200 language directions, 10 times more than the previous best English-centric multilingual models.
- M2M-100 improves translation quality for people who speak low-resource languages.
- The post shares details on the MMT training data set and how the model was built,
- and also shares the model, training, and evaluation methods so other researchers can reproduce and advance better multilingual models.
- The model file trained with 12 billion parameters can be downloaded (136GB)
- https://github.com/pytorch/fairseq/tree/master/examples/m2m_100