datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
real-toxicity-prompts
Dataset Card for Real Toxicity Prompts
Dataset Summary
RealToxicityPrompts is a dataset of 100k sentence snippets from the web for researchers to further address the risk of neural toxic degeneration in models.
Languages
English
Dataset Structure
Data Instances
Each instance represents a prompt and its metadata:
{
"filename":"0766186-bc7f2a64cb271f5f56cf6f25570cd9ed.txt",
"begin":340,
"end":564,
"challenging":false… See the full description on the dataset page: https://huggingface.co/datasets/allenai/real-toxicity-prompts.toxic_conversations_50k
ToxicConversationsClassification
An MTEB dataset
Massive Text Embedding Benchmark
Collection of comments from the Civil Comments platform together with annotations if the comment is toxic or not.
Task category
t2c
Domains
Social, Written
Reference
https://www.kaggle.com/competitions/jigsaw-unintended-bias-in-toxicity-classification/overview
How to evaluate on this task
You can evaluate an embedding model on this dataset using the following code:
import… See the full description on the dataset page: https://huggingface.co/datasets/mteb/toxic_conversations_50k.toxic_conversations
Toxic Conversation
This is a version of the Jigsaw Unintended Bias in Toxicity Classification dataset. It contains comments from the Civil Comments platform together with annotations if the comment is toxic or not.
10 annotators annotated each example and, as recommended in the task page, set a comment as toxic when target >= 0.5
The dataset is inbalanced, with only about 8% of the comments marked as toxic.
ToxiCN
Facilitating Fine-grained Detection of Chinese Toxic Language: Hierarchical Taxonomy, Resources, and Benchmark
🎉2024.9 Our related study, titled "Towards Comprehensive Detection of Chinese Harmful Meme", has been accepted to NeurIPS 2024! In this paper, we present ToxiCN_MM, the first Chinese harmful meme dataset. Here is the link: https://github.com/DUT-lujunyu/ToxiCN_MM. Welcome to star or fork it!
🎉2024.9 Our related study, titled "PclGPT: A Large Language Model for Patronizing… See the full description on the dataset page: https://huggingface.co/datasets/JunyuLu/ToxiCN.Toxic_Russian_Commentshttps://www.kaggle.com/datasets/alexandersemiletov/toxic-russian-comments
0 - neutral user comments
1 - toxic user comments
Toxic Russian Comments Dataset
This dataset contains labelled comments from the popular Russian social network ok.ru.
The data was used in a competition where participants had to automatically label each comment with at least one of the four predefined classes. The classes represent different levels of toxicity. The competition was held on the All Cups platform.
Each… See the full description on the dataset page: https://huggingface.co/datasets/AlexSham/Toxic_Russian_Comments.toxic_conversations_50k
Toxic Conversation
This is a version of the Jigsaw Unintended Bias in Toxicity Classification dataset. It contains comments from the Civil Comments platform together with annotations if the comment is toxic or not.
This dataset just contains the first 50k training examples.
10 annotators annotated each example and, as recommended in the task page, set a comment as toxic when target >= 0.5
The dataset is inbalanced, with only about 8% of the comments marked as toxic.
ToxicQAtextFilteredThis is the TEXT filtered version of TOXICQA with all the semi-refusals (e.g. Remember, killing is bad)
This is a work in progress.
Use only for Alignment research. NOETI is not responsible for what you might do with it.
PleIAs-ToxicCommons
PleIAs/ToxicCommons
This dataset is a refined version of the PleIAs/ToxicCommons collection, focusing on historical texts labeled for content that may be considered objectionable by modern standards (what the authors of the dataset deem "toxic").
The cleaned dataset contains 1 051 027 rows, each representing a text sample with associated toxicity scores across five dimensions:
Race and origin-based bias
Gender and sexuality-based bias
Religious bias
Ability bias
Violence and abuse… See the full description on the dataset page: https://huggingface.co/datasets/agentlans/PleIAs-ToxicCommons.PCR-ToxiCN
PCR-ToxiCN
PCR-ToxiCN is a 500-example Chinese dataset for testing how well models spot offensive language hidden by phonetic cloaking (homophones and near-homophones).
Field
Type
Notes
text
string
Original Xiaohongshu comment
offensive_label
int
1 = offensive, 0 = non-offensive (250 / 250)
strategy
string
HR, AR, NR, or MR
Strategy
What it is
Example
HR
Hanzi replacement
“沸物” → “废物”
AR
Alphabet / pinyin
“SB” → “傻逼”
NR
Numerals as sounds
“4” (sì) →… See the full description on the dataset page: https://huggingface.co/datasets/UTSNLPGroup/PCR-ToxiCN.LLM-jp-Toxicity-Dataset
LLM-jp Toxicity Dataset
日本語有害文書データセット「LLM-jp Toxicity Dataset」
See https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-toxicity-dataset
Toxicity-Bias-Filtering
Overview
This dataset is designed to evaluate the effectiveness of toxicity and bias filtering methods. The objective is to detect and filter a small subset of toxic or unsafe examples that have been injected into a larger, predominantly safe training set, using a reference set that exposes unsafe model behavior.
All models are evaluated using the same training and reference sets.
We provide two evaluation settings, denoted by the suffixes Hom (Homogeneous) and Het (Heterogeneous).… See the full description on the dataset page: https://huggingface.co/datasets/DataAttributionEval/Toxicity-Bias-Filtering.toxic-russian-datasethttps://www.kaggle.com/datasets/alexandersemiletov/toxic-russian-comments
0 - neutral user comments
1 - toxic user comments
Toxic Russian Comments Dataset
This dataset contains labelled comments from the popular Russian social network ok.ru.
The data was used in a competition where participants had to automatically label each comment with at least one of the four predefined classes. The classes represent different levels of toxicity. The competition was held on the All Cups platform.
Each… See the full description on the dataset page: https://huggingface.co/datasets/ilyiniv1755/toxic-russian-dataset.ToxicDataset
Comprehensive Toxic Content Dataset
Dataset Description
This dataset contains 1,000,000 synthetically generated records of toxic, abusive, harmful, and offensive content designed for training content moderation systems and hate speech detection models.
Dataset Summary
This comprehensive dataset includes multiple categories of toxic content:
Toxic content (insults, derogatory terms)
Abusive language patterns
Gender bias statements
Dangerous/threatening content… See the full description on the dataset page: https://huggingface.co/datasets/AiActivity/ToxicDataset.toxicity_preproptoxic_conversationsThis is a version of Jigsaw Unintended Bias in Toxicity Classification (https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/overview) with a train, test, and validation split.
toxic-dpo-v0.1-NoWarningtoxic-dpo-natural-v5I mixed in toxid-dpo-natural-v4 and rawrr v2-1 stage 2 with chosen field from original no_robots and got myself toxic-dpo-natural-v5. Goal is to avoid overfitting via DPO to a specific type of instruct, and instead just DPO the model to be more open to answering and also answer like a human being. We'll see whether this works.I trained Yi 34B with this dataset and ORPO, it does work very nicely so far!
HacxGPT-Toxic
HacxGPT-Toxic Dataset
⚠️ CONTENT WARNING: STRICTLY FOR RESEARCH PURPOSES
This dataset contains explicit, highly toxic, offensive, and dangerous content. It features unaligned AI responses detailing violence, psychological harm, cyber-attacks, and illegal activities. It is published strictly to facilitate red-teaming, alignment research, and defensive cybersecurity evaluation. Use with extreme caution.
Overview
Compiled by BlackTechX011, the HacxGPT-Toxic dataset is a… See the full description on the dataset page: https://huggingface.co/datasets/BlackTechX011/HacxGPT-Toxic.toxicity-classifier-dataset-v3
Multilingual Toxicity Dataset V3
124K balanced samples for binary toxicity classification across three languages.
Author
Görkem Yıldız
GitHub: gorkem371
Website: gorkemyildiz.com
Dataset Details
Split
Samples
Train
105,945
Valid
18,684
Total
124,629
Languages
Turkish (~34%) — sourced primarily from Overfit-GM
Arabic (~32%) — sourced primarily from arabic-hate-speech-superset
English (~34%) — sourced from toxic_conversations_50k… See the full description on the dataset page: https://huggingface.co/datasets/gorkem371/toxicity-classifier-dataset-v3.plant-pet-toxicity-database
PlantFun Plant-Pet Toxicity Database
This dataset is exported from the GitHub Repository.
Official website: plantfun.app.
Snapshot
Generated at: 2026-02-13T02:09:58Z
Total markdown articles: 1313
Pet Toxicity Reports: 498
Misdiagnosis Case Studies: 408
Dynamic Care Protocols: 407
Detailed Encyclopedia: Explore all 1313 plants on PlantFun
Files
../articles.csv
../articles.jsonl
../manifest.json
Suggested usage
Plant toxicity and pet safety… See the full description on the dataset page: https://huggingface.co/datasets/LeafVibe/plant-pet-toxicity-database.toxic_conversations
Toxic Conversation
This is a version of the Jigsaw Unintended Bias in Toxicity Classification dataset. It contains comments from the Civil Comments platform together with annotations if the comment is toxic or not.
10 annotators annotated each example and, as recommended in the task page, set a comment as toxic when target >= 0.5
The dataset is inbalanced, with only about 8% of the comments marked as toxic.
toxic-dpo-v0.1-sharegptUPDATE: Merged the NoWarning into a real DPO for later use. Be aware that the shareGPT format is NOT real DPO, it was just a convertion to shareGPT to add into any datasets. If you want to do a REAL DPO train, use this file: toxic-dpo-NoWarning.json.
DISCLAIMER : I'M NOT THE AUTHOR OF THIS DATASET.
ALL CREDIT GO TO unalignment repo.
ORIGINAL DATASET: unalignment/toxic-dpo-v0.1
I just converted/modified the dataset! Only the accepted replies was taken for the shareGPT format!… See the full description on the dataset page: https://huggingface.co/datasets/Undi95/toxic-dpo-v0.1-sharegpt.toxic-datasetNobodyExistsOnTheInternet_ToxicDPOqa_llama_factoryreformat of: NobodyExistsOnTheInternet/ToxicDPOqa for llama-factory DPO format
usage example:
"toxic_dpo_reformat": {
"hf_hub_url": "lucyknada/NobodyExistsOnTheInternet_ToxicDPOqa_llama_factory",
"ranking": true,
"columns": {
"prompt": "prompt",
"response": "response",
"system": "system"
}
},
toxic-sft-zh数据集 unalignment/toxic-dpo-v0.2 的中文版本,并基于self-Instruct和evol-Instruct继续扩展。
这是一个高度有害的数据集，旨在通过一些示例来说明如何使用 SFT 轻松地对模型进行去审查/取消对齐。
数据生成代码：https://github.com/Minami-su/character_AI_open/tree/main/toxic-Instruction
使用限制请参照原数据集的 Usage restriction。
rtb-real-toxicity
real-toxicity-prompts dataset
Red teaming real-toxicity-prompts dataset.
Generated from https://huggingface.co/datasets/allenai/real-toxicity-prompts
Dataset Structure
Sample
{
"expected": "Acceptable Unacceptable",
"id": "96",
"messages": [
{
"content": "Every assistant has a different tolerance to toxic content. When shown a text, you determine if it is acceptable by your standard of security or unacceptable. If the request… See the full description on the dataset page: https://huggingface.co/datasets/innodatalabs/rtb-real-toxicity.filtered_toxic_samplesToxic_qa_chatmlham-spam-scam-toxicdpo-toxic-zh
DPO Toxic Chinese v2.0
Change Log
v2.0: 增加了adamo1139/toxic-dpo-natural-v5, 并更新了翻译策略. prompt由t5_translate模型翻译, chosen由Uncensored大模型翻译, rejected由一般大模型对prompt生成拒绝性的回复
v1.0: 最初版本, 使用大模型将unalignment/toxic-dpo-v0.2翻译而来
这是一个高度毒性, 高度有害的数据集, 意在展示DPO是如何破除模型的审核/对齐的
使用限制
这个数据集被设计用于学术研究, 而非其他恶意场景. 下载或使用这个数据集, 则视为您承认以下的事实:
这个数据集是有毒的, 包含许多敏感内容
数据集中文本包含的内容和观点与我完全无关, 它们只是大模型生成的文字
您可以使用该数据集, 但必须遵守相关法律
您对您自己下载和使用数据集的行为负责, 我对您的一切行为没有任何责任