Skip to content

Dataset · FredZhang7

Multilingual toxic comments dataset

FredZhang7/toxi-text-3M

Around 3 million real-user texts across 55 languages annotated for toxicity to train and evaluate moderation filters.

What's inside

Contains comments, tweets, and messages written by real people without machine translation, across 55 languages (including Arabic, Spanish, French, Thai, and others). The preprocessed training portion has 2,880,667 rows in CSV format, with 416,529 labeled as toxic and 2,463,773 as neutral.

What you can do with it

  • Training multilingual toxicity classifiers for automated social media and chat moderation
  • Fine-tuning language models and LoRA adapters for filtering unwanted content
  • Evaluating the quality of zero-shot classifiers on low-resource languages
Tasks
Text classification, Zero shot classification
Languages
Arabic, Spanish, Punjabi, Thai, Estonian, French, Finnish, Hungarian and 4 more
Size
1M–10M rows
Downloads this month
711
Likes
32

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("FredZhang7/toxi-text-3M")