README.md

metadata

language:
  - en
license: apache-2.0

• 33.058 texts in total, 14.491 NOT hateful texts and 18.567 HATEFUL texts
• All duplicate values were removed
• Split using sklearn into 80% train and 20% temporary test (stratified label). Then split the test set using 0.50% test and validation (stratified label)
• Split: 80/10/10
• Train set label distribution: 0 ==> 11.593, 1 ==> 14.853
• Validation set label distribution: 0 ==> 1.449, 1 ==> 1.857
• Test set label distribution: 0 ==> 1.449, 1 ==> 1.857
• Combination of 4 publicly available datasets:
• 1. "Ethos" dataset (Mollas et al., 2022)
• 2. HateCheck: Functional Tests for Hate Speech Detection Models (Rottger et al., 2021)
• 3. A Benchmark Dataset for Learning to Intervene in Online Hate Speech (Qian et al., 2019)
• 4. Automated Hate Speech Detection and the Problem of Offensive Language (Davidson, et al., 2017)