Knover: An Overview
Knover is a powerful toolkit designed for generating dialogue grounded in knowledge, and it's built on the PaddlePaddle platform. This tool makes it easier for researchers and developers to efficiently train and execute large-scale dialogue generation models.
Recent Developments
Knover is continually evolving with new features and models. Here are some recent highlights:
- October 2022: Knover introduced Q-TOD, a new system focused on query-driven, task-oriented dialogue.
- March 2022: The toolkit expanded with PLATO-KAG, a novel approach that uses unsupervised learning for end-to-end knowledge-based conversation.
- February 2022: Knover unveiled the TOD-DA dataset, alongside relevant models and code, as part of the DSTC10-Track2 project.
- December 2021: They launched the PLATO-XL dialogue model, boasting up to 11 billion parameters, enhancing dialogue generation capabilities.
- October 2021: AG-DST was introduced, a model designed for flexible dialogue state tracking.
- February 2021: An early implementation from Team 19 was made available in the DSTC9-Track1.
- July 2020: Introduction of PLATO-2, a large-scale generative model featuring a latent space, aimed at open-domain dialogue systems.
Requirements and Installation
To get started with Knover, ensure you meet the following requirements:
- Python version 3.7 or higher.
- PaddlePaddle-GPU version 2.4.0 or higher. Installation instructions can be found on the PaddlePaddle website.
- Your PaddlePaddle version should match your CUDA (recommended version 10.1) and CuDNN (recommended version 7.6) versions. More details are available on the GPU support documentation.
- Additional packages such as
sentencepiece
andtermcolor
. - For distributed training, NCCL is required, downloadable from NCCL's website.
Installation Instructions:
To install Knover locally, you can perform the following:
git clone https://github.com/PaddlePaddle/Knover.git
cd Knover
pip3 install -e .
Alternatively, to setup using PYTHONPATH
:
export PYTHONPATH=/abs/path/to/Knover:$PYTHONPATH
Basic Usage
For detailed instructions on using Knover, refer to the usage documentation.
Disclaimer
The Knover project is intended to foster further research in dialogue generation. It is important to note that Baidu does not bear responsibility for any third-party generations made using the pre-trained system.
Contact Information
For assistance or to report issues while using Knover, the team encourages users to submit an issue on their GitHub page.