Datasets:
dataset_info:
features:
- name: comment_id
dtype: int32
- name: text
dtype: large_string
- name: platform
dtype: int8
- name: annotators
dtype: int64
- name: hate_speech_score
dtype: float64
- name: sentiment
list: int64
- name: respect
list: int64
- name: insult
list: int64
- name: humiliate
list: int64
- name: status
list: int64
- name: dehumanize
list: int64
- name: violence
list: int64
- name: genocide
list: int64
- name: attack_defend
list: int64
- name: hatespeech
list: int64
splits:
- name: train
num_bytes: 9320253
num_examples: 15612
- name: validation
num_bytes: 511148
num_examples: 846
- name: test
num_bytes: 537097
num_examples: 894
download_size: 8699623
dataset_size: 10368498
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
source_datasets:
- ucberkeley-dlab/measuring-hate-speech
license: cc-by-4.0
measuring-hate-speech-votes
Measuring Hate Speech (Kennedy et al., 2020; Sachdeva et al., 2022), one row per comment with vote counts.
The source has one row per (comment, annotator). Each survey item becomes a list of vote counts over its ordinal codes, in code order: 0-4 for the nine Likert items, where a higher code is more hateful (sentiment: strongly positive to strongly negative; respect: strongly respectful to strongly disrespectful; insult, humiliate, dehumanize, violence, genocide: strongly disagree to strongly agree; status: strongly superior to strongly inferior; attack_defend: strongly defending to strongly attacking), and 0-2 for hatespeech (no, unclear, yes). Survey wording: Sachdeva et al. (2022), appendix table 1. Comments with fewer than three annotators are dropped; rows are split 90/5/5 by a hash of the comment id. License: CC BY 4.0, as the original.
Original data: ucberkeley-dlab/measuring-hate-speech. Repackaged as parquet for tasksource by scripts/upload_repackaged.py.