KILT has been built from 11 datasets representing 5 types of tasks:
Fact-checking
Entity linking
Slot filling
Open domain QA
Dialog generation
All these datasets have been grounded in a single pre-processed Wikipedia dump, allowing for fairer and more consistent evaluation as well as enabling new task setups such as multitask and transfer learning with minimal effort. KILT also provides tools to analyze and understand the… See the full description on the dataset page: https://huggingface.co/datasets/facebook/kilt_tasks.