cplane/toxigen-data

---
annotations_creators:
- expert-generated
language_creators:
- machine-generated
multilinguality:
- monolingual
size_categories:
- 100K<n<1M
source_datasets:
- original
task_categories:
- text-classification
task_ids:
- hate-speech-detection
pretty_name: ToxiGen
dataset_info:
- config_name: annotated
features:
- name: text
dtype: string
- name: target_group
dtype: string
- name: factual?
dtype: string
- name: ingroup_effect
dtype: string
- name: lewd
dtype: string
- name: framing
dtype: string
- name: predicted_group
dtype: string
- name: stereotyping
dtype: string
- name: intent
dtype: float64
- name: toxicity_ai
dtype: float64
- name: toxicity_human
dtype: float64
- name: predicted_author
dtype: string
- name: actual_method
dtype: string
splits:
- name: test
num_bytes: 364518
num_examples: 940
- name: train
num_bytes: 3238381
num_examples: 8960
download_size: 768996
dataset_size: 3602899
- config_name: annotations
features:
- name: Input.prompt
dtype: string
- name: Input.text
dtype: string
- name: Input.time
dtype: string
- name: Input.generation_method
dtype: string
- name: Input.prompt_label
dtype: string
- name: Input.target_group
dtype: string
- name: Input.binary_prompt_label
dtype: int64
- name: Answer.annotatorAge
dtype: string
- name: Answer.annotatorGender
dtype: string
- name: Answer.annotatorMinority
dtype: string
- name: Answer.annotatorPolitics.1
dtype: bool
- name: Answer.annotatorPolitics.2
dtype: bool
- name: Answer.annotatorPolitics.3
dtype: bool
- name: Answer.annotatorPolitics.4
dtype: bool
- name: Answer.annotatorPolitics.5
dtype: bool
- name: Answer.annotatorRace
dtype: string
- name: Answer.factSelect
dtype: string
- name: Answer.framingQ
dtype: string
- name: Answer.inGroup.on
dtype: bool
- name: Answer.ingroup.1
dtype: bool
- name: Answer.ingroup.2
dtype: bool
- name: Answer.ingroup.3
dtype: bool
- name: Answer.intent.1
dtype: bool
- name: Answer.intent.2
dtype: bool
- name: Answer.intent.3
dtype: bool
- name: Answer.intent.4
dtype: bool
- name: Answer.intent.5
dtype: bool
- name: Answer.lewd.1
dtype: bool
- name: Answer.lewd.2
dtype: bool
- name: Answer.lewd.3
dtype: bool
- name: Answer.refTarget
dtype: string
- name: Answer.stateFrame
dtype: string
- name: Answer.stateGroup
dtype: string
- name: Answer.stereo.1
dtype: bool
- name: Answer.stereo.2
dtype: bool
- name: Answer.stereo.3
dtype: bool
- name: Answer.toAI.1
dtype: bool
- name: Answer.toAI.2
dtype: bool
- name: Answer.toAI.3
dtype: bool
- name: Answer.toAI.4
dtype: bool
- name: Answer.toAI.5
dtype: bool
- name: Answer.toPER.1
dtype: bool
- name: Answer.toPER.2
dtype: bool
- name: Answer.toPER.3
dtype: bool
- name: Answer.toPER.4
dtype: bool
- name: Answer.toPER.5
dtype: bool
- name: Answer.writer.1
dtype: bool
- name: Answer.writer.2
dtype: bool
- name: HashedWorkerId
dtype: int64
splits:
- name: train
num_bytes: 21933185
num_examples: 27450
download_size: 3350653
dataset_size: 21933185
- config_name: prompts
features:
- name: text
dtype: string
splits:
- name: hate_trans_1k
num_bytes: 585554
num_examples: 1000
- name: neutral_black_1k
num_bytes: 857769
num_examples: 1000
- name: hate_native_american_1k
num_bytes: 480000
num_examples: 1000
- name: neutral_immigrant_1k
num_bytes: 342243
num_examples: 1000
- name: hate_middle_east_1k
num_bytes: 426551
num_examples: 1000
- name: neutral_lgbtq_1k
num_bytes: 914319
num_examples: 1000
- name: neutral_women_1k
num_bytes: 394963
num_examples: 1000
- name: neutral_chinese_1k
num_bytes: 412062
num_examples: 1000
- name: hate_latino_1k
num_bytes: 708000
num_examples: 1000
- name: hate_bisexual_1k
num_bytes: 447794
num_examples: 1000
- name: hate_mexican_1k
num_bytes: 675444
num_examples: 1000
- name: hate_asian_1k
num_bytes: 503093
num_examples: 1000
- name: neutral_mental_disability_1k
num_bytes: 556905
num_examples: 1000
- name: neutral_mexican_1k
num_bytes: 483603
num_examples: 1000
- name: hate_mental_disability_1k
num_bytes: 480620
num_examples: 1000
- name: neutral_bisexual_1k
num_bytes: 915612
num_examples: 1000
- name: neutral_latino_1k
num_bytes: 470000
num_examples: 1000
- name: hate_chinese_1k
num_bytes: 384934
num_examples: 1000
- name: neutral_jewish_1k
num_bytes: 649674
num_examples: 1000
- name: hate_muslim_1k
num_bytes: 425760
num_examples: 1000
- name: neutral_asian_1k
num_bytes: 615895
num_examples: 1000
- name: hate_physical_disability_1k
num_bytes: 413643
num_examples: 1000
- name: hate_jewish_1k
num_bytes: 573538
num_examples: 1000
- name: neutral_muslim_1k
num_bytes: 491659
num_examples: 1000
- name: hate_immigrant_1k
num_bytes: 285309
num_examples: 1000
- name: hate_black_1k
num_bytes: 745295
num_examples: 1000
- name: hate_lgbtq_1k
num_bytes: 577075
num_examples: 1000
- name: hate_women_1k
num_bytes: 389583
num_examples: 1000
- name: neutral_middle_east_1k
num_bytes: 415319
num_examples: 1000
- name: neutral_native_american_1k
num_bytes: 586993
num_examples: 1000
- name: neutral_physical_disability_1k
num_bytes: 458497
num_examples: 1000
download_size: 1698170
dataset_size: 16667706
- config_name: train
features:
- name: prompt
dtype: string
- name: generation
dtype: string
- name: generation_method
dtype: string
- name: group
dtype: string
- name: prompt_label
dtype: int64
- name: roberta_prediction
dtype: float64
splits:
- name: train
num_bytes: 169400442
num_examples: 250951
download_size: 18784380
dataset_size: 169400442
configs:
- config_name: annotated
default: true
data_files:
- split: test
path: annotated/test-*
- split: train
path: annotated/train-*
- config_name: annotations
data_files:
- split: train
path: annotations/train-*
- config_name: prompts
data_files:
- split: hate_trans_1k
path: prompts/hate_trans_1k-*
- split: neutral_black_1k
path: prompts/neutral_black_1k-*
- split: hate_native_american_1k
path: prompts/hate_native_american_1k-*
- split: neutral_immigrant_1k
path: prompts/neutral_immigrant_1k-*
- split: hate_middle_east_1k
path: prompts/hate_middle_east_1k-*
- split: neutral_lgbtq_1k
path: prompts/neutral_lgbtq_1k-*
- split: neutral_women_1k
path: prompts/neutral_women_1k-*
- split: neutral_chinese_1k
path: prompts/neutral_chinese_1k-*
- split: hate_latino_1k
path: prompts/hate_latino_1k-*
- split: hate_bisexual_1k
path: prompts/hate_bisexual_1k-*
- split: hate_mexican_1k
path: prompts/hate_mexican_1k-*
- split: hate_asian_1k
path: prompts/hate_asian_1k-*
- split: neutral_mental_disability_1k
path: prompts/neutral_mental_disability_1k-*
- split: neutral_mexican_1k
path: prompts/neutral_mexican_1k-*
- split: hate_mental_disability_1k
path: prompts/hate_mental_disability_1k-*
- split: neutral_bisexual_1k
path: prompts/neutral_bisexual_1k-*
- split: neutral_latino_1k
path: prompts/neutral_latino_1k-*
- split: hate_chinese_1k
path: prompts/hate_chinese_1k-*
- split: neutral_jewish_1k
path: prompts/neutral_jewish_1k-*
- split: hate_muslim_1k
path: prompts/hate_muslim_1k-*
- split: neutral_asian_1k
path: prompts/neutral_asian_1k-*
- split: hate_physical_disability_1k
path: prompts/hate_physical_disability_1k-*
- split: hate_jewish_1k
path: prompts/hate_jewish_1k-*
- split: neutral_muslim_1k
path: prompts/neutral_muslim_1k-*
- split: hate_immigrant_1k
path: prompts/hate_immigrant_1k-*
- split: hate_black_1k
path: prompts/hate_black_1k-*
- split: hate_lgbtq_1k
path: prompts/hate_lgbtq_1k-*
- split: hate_women_1k
path: prompts/hate_women_1k-*
- split: neutral_middle_east_1k
path: prompts/neutral_middle_east_1k-*
- split: neutral_native_american_1k
path: prompts/neutral_native_american_1k-*
- split: neutral_physical_disability_1k
path: prompts/neutral_physical_disability_1k-*
- config_name: train
data_files:
- split: train
path: train/train-*
---

# Dataset Card for ToxiGen

## Table of Contents
- [Dataset Description](#dataset-description)
- [Dataset Summary](#dataset-summary)
- [Languages](#languages)
- [Dataset Structure](#dataset-structure)
- [Data Fields](#data-instances)
- [Additional Information](#additional-information)
- [Citation Information](#citation-information)

## Sign up for Data Access
To access ToxiGen, first fill out [this form](https://forms.office.com/r/r6VXX8f8vh).

## Dataset Description

- **Repository:** https://github.com/microsoft/toxigen
- **Paper:** https://arxiv.org/abs/2203.09509
- **Point of Contact #1:** [Tom Hartvigsen]([e-mail])
- **Point of Contact #2:** [Saadia Gabriel]([e-mail])

### Dataset Summary

This dataset is for implicit hate speech detection. All instances were generated using GPT-3 and the methods described in [our paper](https://arxiv.org/abs/2203.09509).

### Languages

All text is written in English.

## Dataset Structure

### Data Fields

We release TOXIGEN as a dataframe with the following fields:
- **prompt** is the prompt used for **generation**.
- **generation** is the TOXIGEN generated text.
- **generation_method** denotes whether or not ALICE was used to generate the corresponding generation. If this value is ALICE, then ALICE was used, if it is TopK, then ALICE was not used.
- **prompt_label** is the binary value indicating whether or not the prompt is toxic (1 is toxic, 0 is benign).
- **group** indicates the target group of the prompt.
- **roberta_prediction** is the probability predicted by our corresponding RoBERTa model for each instance.

### Citation Information

```bibtex
@inproceedings{hartvigsen2022toxigen,
title={ToxiGen: A Large-Scale Machine-Generated Dataset for Implicit and Adversarial Hate Speech Detection},
author={Hartvigsen, Thomas and Gabriel, Saadia and Palangi, Hamid and Sap, Maarten and Ray, Dipankar and Kamar, Ece},
booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics},
year={2022}
}
```