license: cc0-1.0
task_categories:
- sentence-similarity
language: - en
size_categories: - 100K<n<1M
pretty_name: ToxicPairs
tags: - not-for-all-audiences
configs:- config_name: raw
data_files:- split: train
path: toxic_pairs_train.parquet - split: test
path: toxic_pairs_test.parquet
- split: train
- config_name: deduplicated
data_files:- split: train
path: toxic_pairs_minhash_train.parquet - split: test
path: toxic_pairs_minhash_test.parquet
- split: train
- config_name: raw
Toxic Pairs (English)
Warning: Contains offensive language, not for all audiences
Because this dataset is marked "not for all audiences", preview isn't available. The columns are:
Index(['sentence1', 'sentence2', 'score', 'label'], dtype='object')
Content marked 'safe' from LlamaGuard3 may still contain profanity/offensive language, though it does not
meet any of the moderation criteria defined for LlamaGuard3. I have set the label for these texts as S0.
Recommended: Use deduplicated subset processed via MinHash for use with training.
Train (raw) has 177549 rows
Test has 10000 rows
Categories are imbalanced:
train_df.label.value_counts()
label
S11 92912
S0 42816
S10 32111
S1 4791
S12 2300
S5 1116
S4 405
S6 305
S8 264
S3 202
S2 179
S7 83
S13 36
S9 29
Name: count, dtype: int64
| Llama Guard 3 categories (S0 for 'safe') | |
|---|---|
| S1: Violent Crimes | S2: Non-Violent Crimes |
| S3: Sex-Related Crimes | S4: Child Sexual Exploitation |
| S5: Defamation | S6: Specialized Advice |
| S7: Privacy | S8: Intellectual Property |
| S9: Indiscriminate Weapons | S10: Hate |
| S11: Suicide & Self-Harm | S12: Sexual Content |
This is a dataset of offensive content mined from several source datasets:
Youtube toxic comments: https://www.kaggle.com/datasets/reihanenamdari/youtube-toxicity-data
Jigsaw: https://www.kaggle.com/datasets/julian3833/jigsaw-toxic-comment-classification-challenge
Cyberbullying Dataset: https://www.kaggle.com/datasets/saurabhshahane/cyberbullying-dataset
Creation Process
I first inferenced the text through LlamaGuard3 and sorted texts into categorical folders based on their llamaguard classification category.
For each classification, I used WordLlama to fuzzy deduplicate the texts, and added them to a BM25s index.
For each text in the classification, I retrieved the top50 results, reranked them using cross-encoder/stsb-roberta-large,
and thresholded them between scores of 0.5 to 0.95. I retained up to 5 results per text.
Finally, I deduplicated them again and stored the results.
Goal
The goal of this dataset is to help embedding models with understanding toxic content for tasks like content moderation,
and benchmarks like ToxicConversationsClassification.