facebookresearch/fairseq — reverse-engineered prompt

Build me a Python research toolkit for training and testing sequence models, similar to Fairseq. I want it to help people work on translation, summarization, language modeling, speech recognition, and general text generation without having to write every training loop from scratch.

It should have a clean command line workflow where I can prepare data, train a model, validate it, generate predictions, and resume from checkpoints. Please include solid defaults for common models like Transformers, CNN and LSTM based models, plus room to add newer research models like wav2vec style speech models or RoBERTa style language models. I also want example projects that show how to run real tasks, not just toy code.

Make the code organized, documented, and testable. Use PyTorch, support GPUs when available, save logs and checkpoints clearly, and make configuration easy to change from the command line. Look up current docs online if you need to.