Knover is a toolkit for knowledge grounded dialogue generation based on PaddlePaddle. Knover allows researchers and developers to carry out efficient training/inference of large-scale dialogue generation models.

• October 2022: We are opening Q-TOD, a novel query-driven task-oriented dialogue system.
• March 2022: We are opening PLATO-KAG, an unsupervised learning approach for end-to-end knowledge-grounded conversation modeling.
• February 2022: We are opening our TOD-DA dataset, models and code in DSTC10-Track2.
• December 2021: We are opening the dialogue generation model of PLATO-XL, with up to 11 billion parameters.
• October 2021: We are opening AG-DST, an amendable generation for dialogue state tracking.
• February 2021: We are opening our implementation (Team 19) in DSTC9-Track1.
• July 2020: We are opening PLATO-2, a large-scale generative model with latent space for open-domain dialogue systems.

• python version >= 3.7
• paddlepaddle-gpu version >= 2.4.0
  • You can install PaddlePaddle following the instructions.
  • The specific version of PaddlePaddle is also based on your CUDA version (recommended version: 10.1) and CuDNN version (recommended version: 7.6). See more information on PaddlePaddle document about GPU support
• sentencepiece
• termcolor
• If you want to run distributed training, you'll also need NCCL
• Install Knover locally:

git clone https://github.com/PaddlePaddle/Knover.git
cd Knover
pip3 install -e .

• Or you can setup PYTHONPATH only:

export PYTHONPATH=/abs/path/to/Knover:$PYTHONPATH

• See usage document.

This project aims to facilitate further research progress in dialogue generation. Baidu is not responsible for the 3rd party's generation with the pre-trained system.

For help or issues using Knover, please submit a GitHub issue.