Published October 19, 2026 | Version v2

MTD-EU-26: Multilingual English–Urdu Text Detoxification Dataset

  • 1. Penn State

Description

This dataset accompanies the study “Multilingual Detoxification of English–Urdu Social-Media Text: A Low-Resource Dataset and LLM Evaluation.”

MTD-EU-26 was developed from 12,045 English and Urdu social-media posts/comments collected from Reddit, X (formerly Twitter), YouTube, and Facebook.

Contents of the public release:

  • English: 9,019 instances from Reddit (2,999), X (3,020), and YouTube (3,000).

  • Urdu: 3,026 instances from Facebook and X.

The publicly released files contain detoxified rewrites and associated research-generated information, where available. The original toxic social-media posts/comments have been excluded because they constitute third-party user-generated content subject to source-platform redistribution restrictions. 

Access to original source text

Original platform post/comment identifiers were not retained during data preprocessing and are therefore unavailable for redistribution or reconstruction of the exact source corpus.

Researchers seeking access to the original toxic social-media text for non-commercial research purposes may contact Dr. Vinayak Elangovan, Program Chair of Computer Science, Penn State Abington (vue9@psu.edu). He will coordinate requests with the corresponding author, Dr. Iqra Ameer (iqa5148@psu.edu). Access is subject to agreement not to redistribute the data or attempt to re-identify users, and to applicable platform terms. 

Intended use

This release supports research on multilingual text detoxification, toxicity mitigation, meaning preservation, and low-resource natural language processing. See the accompanying README (in .zip folder) for information about dataset structure and access restrictions.

Files

MTD-EU-26_v2.0.zip

Files (809.0 kB)

Name Size
md5:7e765d166857e1e7b5d649c1b50f9f52
809.0 kB Preview Download

Additional details

Dates

Submitted
2026-04-14