ParlAI is a unified platform that streamlines the process of researching, training, and evaluating conversational AI models across multiple tasks at once. It provides access to a variety of highly curated, open source datasets to accelerate development, and enables both testing new dialog algorithms and techniques as well as integration of existing models for creating conversational solutions.
ParlAI brings together the community of researchers and developers to push the state of the art in dialog research. It provides a one-stop shop where researchers can submit new tasks and training algorithms to a single, shared repository.
The platform currently supports over 40 public datasets, including popular ones such as SquAD, bAbl tasks, SimpleQuestions, CLEVR, and more. It also integrates with Mechanical Turk and Facebook Messenger for training, evaluation, and more.
ParlAI supports 5 main categories of tasks - question answering, sentence completion (cloze test), goal-oriented dialog, chit-chat dialog, and visual dialog. It also includes examples of training neural models with PyTorch, and can be used with TensorFlow as well.
git clone https://github.com/facebookresearch/ParlAI.git ~/ParlAI cd ~/ParlAI; python setup.py develop
Install PyTorch to work with models that have additional requirements.
Review documentation to familiarize yourself with Agents, Teachers, Worlds, and other concepts in ParlAI.
Explore examples and create a new task.