DETOX
// NLP for Social Good

TextDetox

A multilingual research initiative to automatically transform toxic language into neutral, fluent text — building safer digital spaces across 15 languages worldwide.

Detoxifying Language, One Sentence at a Time

Text detoxification is a Text Style Transfer (TST) task: given a toxic sentence, produce a rewritten version that is non-toxic, meaning-preserving, and fluent. Our research line — pioneered with parallel corpora for English and Russian — has grown into the largest multilingual detoxification ecosystem, covering 14 languages and hosting two international shared tasks at CLEF 2024 & 2025.
☠ Toxic Input
"What the hell are you doing, you absolute idiot? Get out of my sight!"
✦ Detoxified Output
"What are you doing? Please leave."
15
Languages covered — English, Russian, Ukrainian, German, French, Spanish, Italian, Chinese, Japanese, Arabic, Hebrew, Hindi, Hinglish, Tatar, Amharic
3.6k+
Parallel toxic↔neutral sentence pairs in the multilingual dataset, crowd-sourced and human-validated
11
Open-source models on HuggingFace — classifiers, detoxification baselines, and LLM-based evaluators
3
International shared tasks: RUSSE 2022, TextDetox CLEF 2024 & 2025 — with teams from 20+ countries
Supported Languages
🇬🇧 English🇷🇺 Russian 🇺🇦 Ukrainian🇩🇪 German 🇪🇸 Spanish🇫🇷 French 🇮🇹 Italian🇨🇳 Chinese 🇯🇵 Japanese🇸🇦 Arabic 🇮🇱 Hebrew🇮🇳 Hindi🇮🇳 Hinglish Tatar🇪🇹 Amharic

Benchmarking the Community

2025
TextDetox @ CLEF 2025
The second edition of the multilingual detoxification shared task at PAN/CLEF. Extended coverage, updated parallel corpora, and new LLM-based evaluation protocols. Teams compete across 9 languages with both automatic and human judgments.
→ Task Website    → Starter Kit 🤗
2024
TextDetox @ CLEF 2024
First edition of the multilingual detoxification shared task at PAN 2024. Attracted international participants working on English, Russian, Ukrainian, German, Spanish, Chinese, Arabic, Hindi, and Amharic. Benchmark included crowdsourced parallel corpora and human evaluation.
→ Task Website    → Overview Paper
2022
RUSSE Detoxification 2022
The first text detoxification shared task in Russian, held at the Dialogue 2022 conference. Featured the first parallel Russian detoxification corpus and manual human evaluation. Pioneered the crowdsourcing methodology later extended to 14 languages.
→ GitHub Repository

Parallel Corpora for Detoxification

Multilingual · 15 Languages
MultiParaDetox
The flagship multilingual parallel corpus with toxic↔neutral pairs for 15 languages. Collected via crowd-sourcing pipeline extending the original ParaDetox methodology.
3.6k+ pairs15 languagesParallel
→ 🤗 textdetox/multilingual_paradetox
Multilingual · Test Benchmark
MultiParaDetox Test Set
Official held-out test benchmark used across CLEF 2024 & 2025 shared tasks. Human-validated gold references for 9+ languages.
9k samplesMulti-reference
→ 🤗 textdetox/multilingual_paradetox_test
English
ParaDetox (EN)
The first parallel English detoxification dataset. 10,000+ toxic sentences paired with human-written neutral paraphrases. Introduced at ACL 2022 — the foundation that started it all.
10k+ pairsEnglishACL 2022
→ 🤗 s-nlp/paradetox → GitHub
Russian
Russian ParaDetox
The first parallel Russian detoxification corpus, enabling the RUSSE 2022 shared task and the first seq2seq detoxification models for Russian.
RussianParallelRUSSE 2022
→ 🤗 s-nlp/ru_paradetox
Multilingual · Toxicity
Multilingual Toxicity Dataset
Large monolingual toxicity classification dataset across multiple languages. Used for training and fine-tuning the multilingual classifiers.
71.4k samplesClassification
→ 🤗 textdetox/multilingual_toxicity_dataset
Multilingual · XAI
Multilingual Toxic Spans & Lexicon
Fine-grained explainability resources: span-level toxic word annotations and a multilingual lexicon of toxic terms across 14 languages — enabling the explainable detox pipeline.
Toxic Spans: 8.79kLexicon: 176k
→ Toxic Spans → Toxic Lexicon
Multilingual · XAI
Multilingual Toxicity Explained
Human-annotated dataset providing natural language explanations for why a sentence is considered toxic — the first such resource for multiple languages.
8.24k samplesExplanations
→ 🤗 textdetox/multilingual_toxicity_explained
Spanish
ES ParaDetox
Parallel Spanish detoxification corpus collected via the MultiParaDetox pipeline, part of the multilingual expansion effort.
SpanishParallel
→ 🤗 textdetox/es_paradetox

Open-Source Model Zoo

Model Task Size Downloads Link
xlmr-large-toxicity-classifier Classification 0.3B 1.3k 🤗 HF
xlmr-large-toxicity-classifier-v2 Classification 0.6B 934 🤗 HF
roberta_toxicity_classifier Classification 0.3B 49k 🤗 HF
russian_toxicity_classifier Classification 0.3B 7k 🤗 HF
bert-multilingual-toxicity-classifier Classification 0.2B 591 🤗 HF
glot500-toxicity-classifier Classification 0.4B 689 🤗 HF
twitter-xlmr-toxicity-classifier Classification 0.6B 107 🤗 HF
bart-base-detox Generation 0.1B 615 🤗 HF
ruT5-base-detox Generation 0.2B 7 🤗 HF
mbart-detox-baseline Generation 0.6B 9 🤗 HF
mt5-xl-detox-baseline Generation 4B 8 🤗 HF
Llama-pairwise-toxicity-evaluator Evaluation 8B 5 🤗 HF
Llama-pairwise-content-evaluator Evaluation 8B 🤗 HF

How We Measure Quality

🎯
Style Transfer Accuracy
STA
Measures whether the output text is non-toxic. Computed via multilingual toxicity classifiers (XLM-R based). Range: 0 → 1.
📐
Meaning Preservation
SIM
Semantic similarity between toxic input and detoxified output. Computed using multilingual sentence encoders. Ensures no content is lost.
✍️
Fluency
FL
Measures grammaticality and naturalness. Computed as inverse perplexity from a language model. High FL = natural-sounding output.
J = STA × SIM × FL JOINT SCORE — primary ranking metric across all shared task editions
LLM-as-Judge Evaluation (2025)
TextDetox 2025 introduces pairwise LLM-based evaluation using fine-tuned Llama 3 8B models: one judging toxicity style transfer, another judging content preservation — providing human-like assessment at scale. Models are available open-source at textdetox/Llama-pairwise-toxicity-evaluator.

Research Papers

CLEF 2025
Overview of the Multilingual Text Detoxification Task at PAN 2025
Daryna Dementieva, Vitaly Protasov, Nikolay Babakov, Naquee Rizwan, Ilseyar Alimova, Caroline Brun, Vasily Konovalov, Arianna Muti, Chaya Liebeskind, Marina Litvak, Debora Nozza, Shehryaar Shah Khan, Sotaro Takeshita, Natalia Vanetik, Abinew Ali Ayele, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Ashraf Elnagar, Animesh Mukherjee, and Alexander Panchenko
2025 edition of the Shared task with new languages: French, Italian, Hebrew, Hinglish, Japanese, and Tatar.
COLING 2025
Multilingual and Explainable Text Detoxification with Parallel Corpora
Daryna Dementieva, Nikolay Babakov, Amit Ronen, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Daniil Moskovskiy, Elisei Stakovskii, Eran Kaufman, Ashraf Elnagar, Animesh Mukherjee, Alexander Panchenko
Extended multilingual corpora (DE, ZH, AR, HI, AM) + explainable analysis + Chain-of-Thought detox prompting
CLEF Working Notes 2024
Overview of the Multilingual Text Detoxification Task at PAN 2024
Daryna Dementieva, Daniil Moskovskiy, Nikolay Babakov, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Dmitry Ustalov, Elisei Stakovskii, Alisa Smirnova, Ashraf Elnagar, Animesh Mukherjee, Alexander Panchenko
Comprehensive overview of TextDetox CLEF2024: participants, systems, results, and findings
NAACL 2024
MultiParaDetox: Extending Text Detoxification with Parallel Data to New Languages
Daryna Dementieva, Nikolay Babakov, Alexander Panchenko
First automated multilingual parallel corpus collection pipeline; state-of-the-art detox in 9 languages
IJCNLP-AACL 2023
Exploring Methods for Cross-lingual Text Style Transfer: The Case of Text Detoxification
Daryna Dementieva, Daniil Moskovskiy, David Dale, Alexander Panchenko
First study of simultaneous translation + detoxification; new automatic evaluation metrics with higher human correlation
ACL 2022
ParaDetox: Detoxification with Parallel Data
Varvara Logacheva*, Daryna Dementieva*, Sergey Ustyantsev, Daniil Moskovskiy, David Dale, Irina Krotova, Nikita Semenov, Alexander Panchenko (* equal contribution)
First parallel English detoxification dataset. 10k+ crowdsourced pairs. 98+ citations. The paper that started the field.
ACL Student Workshop 2022
Exploring Cross-lingual Text Detoxification with Large Multilingual Language Models
Daniil Moskovskiy, Daryna Dementieva, Alexander Panchenko
First investigation of multilingual and cross-lingual detoxification behavior in large pretrained models
Dialogue 2022
RUSSE-2022: Findings of the First Russian Detoxification Shared Task Based on Parallel Corpora
Daryna Dementieva, Varvara Logacheva, Irina Nikishina, Alena Fenogenova, David Dale, Irina Krotova, Nikita Semenov, Tatiana Shavrina, Alexander Panchenko
First Russian detox shared task; analysis of automatic vs. human evaluation; crowdsourcing pipeline methodology
Multimodal Technologies & Interaction 2021
Methods for Detoxification of Texts for the Russian Language
Daryna Dementieva, Daniil Moskovskiy, Varvara Logacheva, David Dale, Olga Kozlova, Nikita Semenov, Alexander Panchenko
First study of automatic Russian text detoxification using BERT-based editing and GPT-2 seq2seq approaches. 81+ citations.

Who We Are

The TextDetox initiative is led by Daryna Dementieva (postdoctoral researcher at TU Munich), in collaboration with an international team of NLP researchers from TU Munich, Skoltech, IIT Kharagpur, University of Hamburg, UAE University, Bar-Ilan University, and others.

All contributors: Daryna Dementieva, Nikolay Babakov, Vitaly Protasov, Elisei Stakovskii, Debora Nozza, Caroline Brun, Chaya Liebeskind, Arianna Muti, Sotaro Takeshita, Alisa Smirnova, Daniil Moskovskiy, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Abinew Ali Ayele, Dmitry Ustalov, Ashraf Elnagar, Animesh Mukherjee, Alexander Panchenko.

We are happy to extend our research to more languages, cultures, and dimensions! 🌍 Contact: Daryna Dementieva on HuggingFace · TUM Profile · ACL Anthology

🤗 HuggingFace Org ⚙ ParaDetox GitHub ⚙ RUSSE 2022 GitHub 🏆 Join CLEF 2025 📰 Press Coverage 🎬 PyData Berlin Talk