--- license: apache-2.0 language: - no - es - so - ca - af - it - nl - hi - cy - ar - sv - cs - pl - de - lt - sq - uk - tl - sl - hr - en - fi - vi - id - da - ko - bg - mr - ja - bn - ro - pt - fr - hu - tr - zh - mk - ur - sk - ne - et - sw - ru - multilingual task_categories: - text-classification - zero-shot-classification tags: - nlp - moderation size_categories: - 10K<n<100K --- This is a large corpus of 42,619 preprocessed text messages and emails sent by humans in 43 languages. `is_spam=1` means spam and `is_spam=0` means ham. 1040 rows of balanced data, consisting of casual conversations and scam emails in ≈10 languages, were manually collected and annotated by me, with some help from ChatGPT. <br> ### Some preprcoessing algorithms - [spam_assassin.js](./spam_assassin.js), followed by [spam_assassin.py](./spam_assassin.py) - [enron_spam.py](./enron_spam.py) <br> ### Data composition  <br> ### Description To make the text format between sms messages and emails consistent, email subjects and content are separated by two newlines: ```python text = email.subject + "\n\n" + email.content ``` <br> ### Suggestions - If you plan to train a model based on this dataset alone, I recommend adding **some** rows with `is_toxic=0` from `FredZhang7/toxi-text-3M`. Make sure the rows aren't spam. <br> ### Other Sources - https://huggingface.co/datasets/sms_spam - https://github.com/MWiechmann/enron_spam_data - https://github.com/stdlib-js/datasets-spam-assassin - https://repository.ortolang.fr/api/content/comere/v3.3/cmr-simuligne.html