A unified platform for sharing, training and evaluating dialogue models across many tasks.

Many popular datasets available all in one place -- from open-domain chitchat to visual question answering.

A wide set of reference models -- from retrieval baselines to Transformers.

Seamless integration of Amazon Mechanical Turk for data collection, training and human evaluation.

Get Started Fork me on GitHub

What's New

2019-09-16: The Dialogue Safety project page is updated with pretrained models, data, and an interactive demonstration.

2019-06-12: The Personality-Captions project page is updated with pretrained models, specialized model code, and an interactive demo.

2019-06-12: Added HotpotQA as another dataset.

2019-05-15: Added QuAC, Question Answering in Context dataset, and COQA, Conversational Question Answering Challenge datasets.

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]

2019-03-06: Added Self-feeding Chatbot for leveraging user textual feedback to improve the chatbot's abilities. [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!

2018-03-05: 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!

Get Started

Check out our GitHub repository:

Run this command:
git clone
cd ParlAI; python develop


Display 10 random examples from task 1 of the "1k training examples" bAbI task:

Run this command:
python examples/ -t babi:task1k:1

Displays 100 random examples from multitasking on the bAbI task and the SQuAD dataset at the same time:

Run this command:
python examples/ -t babi:task1k:1,squad -n 100

Evaluate an IR baseline model on the validation set of the Movies Subreddit dataset:

Run this command:
python examples/ -m ir_baseline -t "#moviedd-reddit" -dt valid

Display the predictions of that same IR baseline model:

Run this command:
python examples/ -m ir_baseline -t "#moviedd-reddit" -dt valid

Trains an attentive LSTM model on the SQuAD dataset with a batch size of 32 examples (pytorch and regex):

Run this command:
python examples/ -m drqa -t squad -bs 32 -mf /tmp/model_drqa

For more examples, please read our tutorial. To learn more about ParlAI, click here.