facebookresearch/fairseq-lua

Fairseq-lua is a machine translation toolkit from Facebook AI Research, written in Lua and built on top of the Torch framework. It was designed to train and run neural network models that translate text from one language to another. The toolkit implements specific model architectures based on convolutional neural networks rather than the attention-based designs that became more common later, and it comes with pre-trained models for translating English to French, German, and Romanian. Training a new model requires a computer with an NVIDIA GPU. Running inference (translating with an already trained model) works on CPU as well. The installation is done through LuaRocks, a Lua package manager, and the toolkit includes command-line tools for preprocessing text, training models, and generating translations. The README opens with a prominent note that this version is no longer actively developed. A newer Python and PyTorch version called fairseq-py became the focus of new work, and the Lua version is preserved for reference but is provided without support. Anyone starting a new translation project would likely use the newer Python version instead. The project was published alongside two academic papers describing the convolutional model architectures it implements, and it includes citation instructions for researchers who use the code.