mlsanityData to 7 Oct 2026

Exploring the Limits of Domain-Adaptive Training for Detoxifying Large-Scale Language Models

arXiv 2202.04173 ↗First posted 8 Feb 2022cs.CL cs.AI cs.CYcited by 67

Not on the map. Why?

Abstract (arXiv)

Pre-trained language models (LMs) are shown to easily generate toxic language. In this work, we systematically explore domain-adaptive training to reduce the toxicity of language models. We conduct this study on three dimensions: training corpus, model size, and parameter efficiency. For the training corpus, we propose to leverage the generative power of LMs and generate nontoxic datasets for domain-adaptive training, which mitigates the exposure bias and is shown to be more data-efficient than using a curated pre-training corpus. We demonstrate that the self-generation method consistently outperforms the existing baselines across various model sizes on both automatic and human evaluations, even when it uses a 1/3 smaller training corpus. We then comprehensively study detoxifying LMs with parameter sizes ranging from 126M up to 530B (3x larger than GPT-3), a scale that has never been studied before. We find that i) large LMs have similar toxicity levels as smaller ones given the same pre-training corpus, and ii) large LMs require more endeavor to detoxify. We also explore parameter-efficient training methods for detoxification. We demonstrate that adding and training adapter-only layers in LMs not only saves a lot of parameters but also achieves a better trade-off between toxicity and perplexity than whole model adaptation for the large-scale models.

Key facts. Cited by 67 papers in the map (data through 7 Oct 2026). Builds on 45 of its 65 references that are in the map. First posted on arXiv 8 Feb 2022.

Builds on

45 of 65 references.

References · 45

  1. 2017Attention Is All You Needcited by 59,932
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  4. 2021Zero-Shot Text-to-Image Generationcited by 4,276
  5. 2019Parameter-Efficient Transfer Learning for NLPcited by 4,195
  6. 2019HellaSwag: Can a Machine Really Finish Your Sentence?cited by 4,018
  7. 2021Prefix-Tuning: Optimizing Continuous Prompts for Generationcited by 3,878
  8. 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,769
  9. 2021Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsitycited by 3,014
  10. 2019The Curious Case of Neural Text Degenerationcited by 2,881
  11. 2019PIQA: Reasoning about Physical Commonsense in Natural Languagecited by 2,818
  12. 2021Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processingcited by 2,222
33 more
  1. 2020Extracting Training Data from Large Language Modelscited by 2,144
  2. 2019BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questionscited by 2,140
  3. 2019Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelismcited by 1,829
  4. 2020Don't Stop Pretraining: Adapt Language Models to Domains and Taskscited by 1,548
  5. 2021Calibrate Before Use: Improving Few-Shot Performance of Language Modelscited by 1,371
  6. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  7. 2015A Diversity-Promoting Objective Function for Neural Conversation Modelscited by 1,068
  8. 2022Red Teaming Language Models with Language Modelscited by 983
  9. 2016The LAMBADA dataset: Word prediction requiring a broad discourse contextcited by 960
  10. 2020Exploiting Cloze Questions for Few Shot Text Classification and Natural Language Inferencecited by 934
  11. 2017RACE: Large-scale ReAding Comprehension Dataset From Examinationscited by 851
  12. 2019Adversarial NLI: A New Benchmark for Natural Language Understandingcited by 798
  13. 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 793
  14. 2015Scheduled Sampling for Sequence Prediction with Recurrent Neural Networkscited by 779
  15. 2016Sequence-Level Knowledge Distillationcited by 761
  16. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 709
  17. 2020It's Not Just Size That Matters: Small Language Models Are Also Few-Shot Learnerscited by 668
  18. 2022Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Modelcited by 557
  19. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 434
  20. 2021BOLD: Dataset and Metrics for Measuring Biases in Open-Ended Language Generationcited by 410
  21. 2019On Measuring Social Biases in Sentence Encoderscited by 397
  22. 2020GeDi: Generative Discriminator Guided Sequence Generationcited by 377
  23. 2018WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representationscited by 370
  24. 2021Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLPcited by 336
  25. 2020AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Promptscited by 246
  26. 2021Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Modelscited by 227
  27. 2019Gender Bias in Contextualized Word Embeddingscited by 203
  28. 2021Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasetscited by 179
  29. 2021Challenges in Detoxifying Language Modelscited by 169
  30. 2021Detoxifying Language Models Risks Marginalizing Minority Voicescited by 106
  31. 2019Evaluating the Underlying Gender Bias in Contextualized Word Embeddingscited by 94
  32. 2020The Radicalization Risks of GPT-3 and Advanced Neural Language Modelscited by 76
  33. 2021Just Say No: Analyzing the Stance of Neural Dialogue Generation in Offensive Contextscited by 56

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 4; 2023: 20; 2024: 25; 2025: 14; 2026: 4.

67 papers.

Sort:

By area:

  1. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,565
  2. 2023GPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Ciphercited by 401
  3. 2023Pretraining Language Models with Human Preferencescited by 245
  4. 2023A Wolf in Sheep's Clothing: Generalized Nested Jailbreak Prompts can Fool Large Language Models Easilycited by 242
  5. 2023In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steeringcited by 240
  6. 2022Factuality Enhanced Language Models for Open-Ended Text Generationcited by 200
  7. 2025GuardReasoner: Towards Reasoning-based LLM Safeguardscited by 89
  8. 2022Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Surveycited by 80
  9. 2024BadChain: Backdoor Chain-of-Thought Prompting for Large Language Modelscited by 74
  10. 2024FlipAttack: Jailbreak LLMs via Flippingcited by 73
  11. 2024Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systemscited by 63
  12. 2025Safety in Large Reasoning Models: A Surveycited by 58
38 more
  1. 2024Fairness in Large Language Models: A Taxonomic Surveycited by 57
  2. 2023All Languages Matter: On the Multilingual Safety of Large Language Modelscited by 51
  3. 2023Self-Detoxifying Language Models via Toxification Reversalcited by 45
  4. 2023On the Challenges of Using Black-Box APIs for Toxicity Evaluation in Researchcited by 44
  5. 2024Stress-Testing Capability Elicitation With Password-Locked Modelscited by 42
  6. 2023Click: Controllable Text Generation with Sequence Likelihood Contrastive Learningcited by 41
  7. 2023Towards Safe Self-Distillation of Internet-Scale Text-to-Image Diffusion Modelscited by 41
  8. 2024Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicitycited by 33
  9. 2022Unified Detoxifying and Debiasing in Language Generation via Inference-time Adaptive Optimizationcited by 32
  10. 2024Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMscited by 32
  11. 2023ToViLaG: Your Visual-Language Generative Model is Also An Evildoercited by 27
  12. 2023Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Modelscited by 23
  13. 2023A Pseudo-Semantic Loss for Autoregressive Models with Logical Constraintscited by 21
  14. 2024Parameter-Efficient Detoxification with Contrastive Decodingcited by 18
  15. 2024From One to Many: Expanding the Scope of Toxicity Mitigation in Language Modelscited by 18
  16. 2024Walking in Others' Shoes: How Perspective-Taking Guides Large Language Models in Reducing Toxicity and Biascited by 17
  17. 2023Adding Instructions during Pretraining: Effective Way of Controlling Toxicity in Language Modelscited by 16
  18. 2024Efficient LLM Jailbreak via Adaptive Dense-to-sparse Constrained Optimizationcited by 16
  19. 2023Harnessing the Power of LLMs: Evaluating Human-AI Text Co-Creation through the Lens of News Headline Generationcited by 14
  20. 2022Enabling Classifiers to Make Judgements Explicitly Aligned with Human Valuescited by 13
  21. 2024Controllable Generation via Locally Constrained Resamplingcited by 11
  22. 2024RSA-Control: A Pragmatics-Grounded Lightweight Controllable Text Generation Frameworkcited by 11
  23. 2024SATA: A Paradigm for LLM Jailbreak via Simple Assistive Task Linkagecited by 10
  24. 2023Unpacking the Ethical Value Alignment in Big Modelscited by 9
  25. 2023Transformation vs Tradition: Artificial General Intelligence (AGI) for Arts and Humanitiescited by 9
  26. 2024DESTEIN: Navigating Detoxification of Language Models via Universal Steering Pairs and Head-wise Activation Fusioncited by 9
  27. 2025Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editingcited by 9
  28. 2023BiasTestGPT: Using ChatGPT for Social Bias Testing of Language Modelscited by 8
  29. 2025UniDetox: Universal Detoxification of Large Language Models via Dataset Distillationcited by 8
  30. 2025Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoderscited by 8
  31. 2024Jailbreaking LLMs with Arabic Transliteration and Arabizicited by 7
  32. 2024Attribute Controlled Fine-tuning for Large Language Models: A Case Study on Detoxificationcited by 7
  33. 2025IF-GUIDE: Influence Function-Guided Detoxification of LLMscited by 7
  34. 2023CMD: a framework for Context-aware Model self-Detoxificationcited by 6
  35. 2025BiasFilter: An Inference-Time Debiasing Framework for Large Language Modelscited by 6
  36. 2024FrenchToxicityPrompts: a Large Benchmark for Evaluating and Mitigating Toxicity in French Textscited by 5
  37. 2025Evaluating and Improving Robustness in Large Language Models: A Survey and Future Directionscited by 5
  38. 2024Composable Interventions for Language Modelscited by 4

50 most cited of 67 All 67 (JSON)