mlsanityData to 7 Oct 2026

Mitigating harm in language models with conditional-likelihood filtration

arXiv 2108.07790 ↗First posted 4 Aug 2021cs.CL cs.LGcited by 38

Not on the map. Why?

Abstract (arXiv)

Language models trained on large-scale unfiltered datasets curated from the open web acquire systemic biases, prejudices, and harmful views from their training data. We present a methodology for programmatically identifying and removing harmful text from web-scale datasets. A pretrained language model is used to calculate the log-likelihood of researcher-written trigger phrases conditioned on a specific document, which is used to identify and filter documents from the dataset. We demonstrate that models trained on this filtered dataset exhibit lower propensity to generate harmful text, with a marginal decrease in performance on standard language modeling benchmarks compared to unfiltered baselines. We provide a partial explanation for this performance gap by surfacing examples of hate speech and other undesirable content from standard language modeling benchmarks. Finally, we discuss the generalization of this method and how trigger phrases which reflect specific values can be used by researchers to build language models which are more closely aligned with their values.

Key facts. Cited by 38 papers in the map (data through 7 Oct 2026). Builds on 10 of its 16 references that are in the map. First posted on arXiv 4 Aug 2021.

Builds on

10 of 16 references.

References · 10

  1. 2020Language Models are Few-Shot Learnerscited by 32,199
  2. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  3. 2016Pointer Sentinel Mixture Modelscited by 2,992
  4. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  5. 2016The LAMBADA dataset: Word prediction requiring a broad discourse contextcited by 960
  6. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 709
  7. 2021Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLPcited by 336
  8. 2013One Billion Word Benchmark for Measuring Progress in Statistical Language Modelingcited by 302
  9. 2021Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasetscited by 179
  10. 2021Censorship of Online Encyclopedias: Implications for NLP Modelscited by 5

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 2; 2022: 5; 2023: 12; 2024: 11; 2025: 4; 2026: 4.

38 papers.

Sort:

By area:

  1. 2022Training language models to follow instructions with human feedbackcited by 15,712
  2. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  3. 2023Bias and Fairness in Large Language Models: A Surveycited by 626
  4. 2022Using DeepSpeed and Megatron to Train Megatron-Turing NLG 530B, A Large-Scale Generative Language Modelcited by 557
  5. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 488
  6. 2023Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignmentcited by 394
  7. 2023Low-Resource Languages Jailbreak GPT-4cited by 290
  8. 2023Aligning Language Models with Preferences through f-divergence Minimizationcited by 118
  9. 2025A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deploymentcited by 110
  10. 2024Secret Collusion among AI Agents: Multi-Agent Deception via Steganographycited by 107
  11. 2024Bias and Unfairness in Information Retrieval Systems: New Challenges in the LLM Eracited by 97
  12. 2023A Survey on Fairness in Large Language Modelscited by 91
26 more
  1. 2022Language Generation Models Can Cause Harm: So What Can We Do About It? An Actionable Surveycited by 80
  2. 2023Attack Prompt Generation for Red Teaming and Defending Large Language Modelscited by 73
  3. 2024Fairness in Large Language Models: A Taxonomic Surveycited by 57
  4. 2023Using In-Context Learning to Improve Dialogue Safetycited by 41
  5. 2023Beyond the Safeguards: Exploring the Security Risks of ChatGPTcited by 36
  6. 2024Navigating LLM Ethics: Advancements, Challenges, and Future Directionscited by 28
  7. 2022From plane crashes to algorithmic harm: applicability of safety engineering frameworks for responsible MLcited by 26
  8. 2023Adding Instructions during Pretraining: Effective Way of Controlling Toxicity in Language Modelscited by 16
  9. 2023A Security Risk Taxonomy for Prompt-Based Interaction With Large Language Modelscited by 12
  10. 2023Tackling Bias in Pre-trained Language Models: Current Trends and Under-represented Societiescited by 11
  11. 2026Shaping capabilities with token-level data filteringcited by 11
  12. 2024Detectors for Safe and Reliable LLMs: Implementations, Uses, and Limitationscited by 10
  13. 2023Transformation vs Tradition: Artificial General Intelligence (AGI) for Arts and Humanitiescited by 9
  14. 2025IF-GUIDE: Influence Function-Guided Detoxification of LLMscited by 7
  15. 2024Tiny Refinements Elicit Resilience: Toward Efficient Prefix-Model Against LLM Red-Teamingcited by 6
  16. 2024Unlearning Climate Misinformation in Large Language Modelscited by 4
  17. 2025Paladin: Defending LLM-enabled Phishing Emails with a New Trigger-Tag Paradigmcited by 4
  18. 2022DialGuide: Aligning Dialogue Model Behavior with Developer Guidelinescited by 3
  19. 2024Is Generative AI the Next Tactical Cyber Weapon For Threat Actors? Unforeseen Implications of AI Generated Cyber Attackscited by 2
  20. 2024Nevermind: Instruction Override and Moderation in Large Language Modelscited by 1
  21. 2024Aligning language models with human preferencescited by 1
  22. 2026Bye Bye Perspective API: Lessons for Building and Governing Measurement Infrastructurecited by 1
  23. 2021No News is Good News: A Critique of the One Billion Word Benchmarkcited by 0
  24. 2025Textual Data Bias Detection and Mitigation -- An Extensible Pipeline with Experimental Evaluationcited by 0
  25. 2026Harder to Defend: Towards Chinese Toxicity Attacks via Implicit Enhancement and Obfuscation Rewritingcited by 0
  26. 2026When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigationcited by 0