facebookresearch/parlai

ParlAI, pronounced "par-lay", is a Python framework from Facebook Research for building, training, and testing AI systems that carry on conversations. The framework is aimed at researchers who study dialogue: systems that answer questions, maintain chitchat, complete tasks through conversation, or respond to images. One of its main offerings is a unified collection of over 100 publicly available conversation datasets, all accessible through the same code interface. Instead of spending time finding each dataset and writing custom loading code, you can pull from a large library of research datasets including question-answering sets, open-domain chat corpora, and visual question answering collections. The framework also includes a set of pre-built baseline models and a collection of pre-trained models you can load and run without training anything yourself. Beyond datasets and models, ParlAI supports collecting new training data from human workers through Amazon Mechanical Turk, and connecting conversation agents to real users through Facebook Messenger. This makes it possible to go from training and evaluation in research settings to actual human interaction without switching tools. The framework requires Python 3.8 or higher and runs on Linux or macOS. Windows is not officially supported, though the README notes that some users have had success with it. Installation is available through pip. Facebook Research created and maintains the project. It was described in a 2017 academic paper titled "ParlAI: A Dialog Research Software Platform" and has continued to grow since. The project website at parl.ai has additional documentation and tutorials, and an interactive notebook-based tutorial is available for anyone who wants to try it without setting up a local environment.