# Dataset card for dynamically generated dataset hate speech detection

## Dataset summary

This dataset that was dynamically generated for training and improving hate speech detection models. A group of trained annotators generated and labeled challenging examples so that hate speech models could be tricked and consequently improved. This dataset contains about 40,000 examples of which 54% are labeled as hate speech. It also provides the target of hate speech, including vulnerable, marginalized, and discriminated groups. Overall, this is a balanced dataset which makes it different from the already available hate speech datasets you can find on the web.

This dataset was presented in the article [Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection published](https://aclanthology.org/2021.acl-long.132.pdf) in 2021. The article describes the process for generating and annotating the data. Also, it describes how they used the generated data for training and improving hate speech detection models. The full author list is the following: Bertie Vidgen (The Alan Turing Institute), Tristan Thrush (Facebook), Zeerak Waseem (University of Sheffield), and Douwe Kiela (Facebook).