Auditing the Helpful and Harmless Dataset: Identifying Limitations and Potential Harms in Safety Mitigation for Large Language Models
While widely used for mitigating harms in large language models, the Helpful and Harmless (HH) dataset exhibits significant shortcomings, potentially leading to exaggerated safety behaviors and perpetuating harmful associations with demographic groups.
Assessing and Improving the Ability of Large Language Models to Self-Correct Harmful Content Generation
This research paper introduces a novel approach to enhance the safety of large language models (LLMs) by developing a benchmark to evaluate and improve their ability to self-correct, or "course-correct," during the generation of harmful content.
Evaluating the Safety of Large Language Models in Chinese Contexts: The ChineseSafe Benchmark
This research paper introduces ChineseSafe, a comprehensive benchmark dataset designed to evaluate the safety of large language models (LLMs) specifically in handling Chinese content, addressing the limitations of existing benchmarks by including categories like political sensitivity, pornography, and variant/homophonic words.
DEAN: A Training-Free Method for Mitigating Fairness and Privacy Conflicts in Large Language Models
Deactivating specific neurons in large language models (LLMs) can mitigate the trade-off between fairness and privacy awareness, leading to simultaneous improvements in both areas without compromising general capabilities.
Enhancing Large Language Model Safety: A Framework for Identifying and Explaining Legal Implications Using Prompt Engineering and Knowledge Graphs
This paper proposes a framework that combines prompt engineering and legal knowledge graphs to improve the safety of large language models (LLMs) by identifying and explaining potential legal implications of LLM-generated recommendations.
On the Role of Specific Attention Heads in Large Language Model Safety and How They Can Be Exploited
This research paper presents evidence that specific attention heads within large language models (LLMs) play a crucial role in safety mechanisms, and that ablating these heads can significantly compromise the model's ability to reject harmful queries while having minimal impact on helpfulness.
대규모 언어 모델에서 과생성 및 선호도 최적화를 통해 안전성을 개선하고 과도한 거부를 줄이는 POROver
대규모 언어 모델의 안전성과 유용성 사이의 균형을 맞추는 것은 매우 중요하며, 고급 교사 모델을 사용한 훈련 데이터 과생성과 선호도 최적화 기법을 통해 안전성을 유지하면서 과도한 거부를 줄일 수 있습니다.
Improving Safety and Reducing Overrefusal in Large Language Models Through Overgeneration and Preference Optimization
Overgeneration of training data using advanced language models, combined with preference optimization techniques, can effectively improve the safety and reduce overrefusal in large language models, leading to a better balance between safety and usefulness.
PKU-SafeRLHF: A Large-Scale Dataset for Enhancing the Safety of Large Language Models
The PKU-SafeRLHF dataset aids in aligning large language models with human preferences for safety and helpfulness by providing a large-scale resource for training and evaluating safety alignment techniques.
Mitigating Safety Risks in Fine-tuned Large Language Models: The SafetyLock Approach
Fine-tuning large language models, even on seemingly benign datasets, can introduce significant safety risks, and SafetyLock offers a novel, efficient, and transferable solution to mitigate these risks by identifying and steering specific safety-sensitive attention heads within the model's architecture.
© 2024 by Linnk AI