2019-05-03: The What makes a good Conversation project page is now available with pretrained models.
2019-03-15: The Wizard of Wikipedia project page is updated with pretrained models and specialized model code.
2019-03-09: Added LIGHT text adventure game research platform for learning to speak and act. [press]
2019-02-07: Added BERT Ranker agents, several variations of a ranking model based on the pretrained language model BERT.
2019-01-16: ParlAI has been relicensed under the MIT open source license.
2018-12-13: Added Daily Dialog, an open-domain daily dialogue dataset.
2018-11-05: Added Wizard of Wikipedia, a dataset for knowledge-powered conversation.
2018-11-02: Added Image-Chat, a dataset for engaging personality-conditioned dialogue grounded in images.
2018-10-25: Added Personality-Captions, a dataset for engaging image captioning via personality.
2018-08-29: Added new cleaner version of seq2seq model with new TorchAgent parent class, along with folder (parlai/legacy_agents) for deprecated model code
2018-07-17: Added Qangaroo (a.k.a. WikiHop and MedHop), two reading comprehension datasets with multiple hops, and SQuAD 2.
2018-05-22: Two new tasks added: COCO Image Captioning and Flickr30k Entities
2018-04-13: NIPS ConvAI2 competition! Train Dialogue Agents to chat about personal interests and get to know their dialogue partner -- using the PersonaChat dataset as a training source, with data and baseline code in ParlAI. Competition starts now! Ends September 1st.
2018-03-13: Added ParlAI-Messenger, a new method for connecting human agents to a world in ParlAI using Facebook Messenger. Host your bots on Facebook Messenger to expose them to a broad audience!
2018-03-07: Added IBM's sequence to sequence model to parlai/agents. To use it, just set --model ibm_seq2seq.
2018-03-05: Added Multimodal Low-Rank Bilinear Attention Network (MLB) model for VQA V1 and V2 tasks, adapted from an implementation here based on this paper. To use it, please follow the instructions in the agent file.
2018-02-12: Added a Wikipedia task, which provides a dump of Wikipedia articles from 2/3/2018.
2018-02-07: Added a language model adapted from this Pytorch model to parlai/agents.
2018-01-23: Several new tasks added: SNLI, MultiNLI, COPA, NarrativeQA, Twitter and Persona-Chat.
2017-12-14: Fast, multiprocessed data loading supported with Pytorch data loader
2017-11-30: Several new tasks added: SCAN, ConvAI, NVLR and ISWLT14.
2017-10-19: ParlAI Request For Proposals: Winners Announced!
2017-10-13: New model added: Fairseq-py
2017-10-12: New task added: Stanford's MutualFriends
2017-09-22: New task added: babi+
2017-09-21: New task added: WMT En-De training set, with more WMT tasks on the way
2017-08-25: New task added: Deal or No Deal
2017-08-15: New task added: CLEVR
2017-07-20: ParlAI Request For Proposals: Funding university teams - 7 awards are available - deadline Aug 25
2017-07-20: added building an (seq2seq) agent tutorial
2017-07-12: Several new tasks added: MS Marco, TriviaQA, InsuranceQA, personalized-dialog and MNIST_QA
2017-06-27: ExecutableWorld class for interactive worlds with dialog
2017-06-21: MTurk now supports multiple assignments per HIT
2017-06-20: updated MTurk tutorial to reflect new design
2017-06-20: MTurk now uses general world and agent classes
2017-06-16: added Creating a New Task tutorial
2017-05-31: added Seq2Seq model
2017-05-30: added interactive mode with local human agent
2017-05-22: added MTurk tutorial
2017-05-14: added basic tutorial
2017-05-15: ParlAI press: TechCrunch, CNBC, The Verge, Scientific American, Engadget, Venture Beat, Wired, MIT Technology review.
2017-05-12: added VQA V2.0 and Visual Dialog V0.9 tasks
2017-05-01: ParlAI released!