AI BriefingLLM Reliability Evaluation Leaderboard Released
Key point
A new Safety Leaderboard based on DecodingTrust, which comprehensively evaluates LLMs' toxicity, bias, security, and more, has been released.
Details
To evaluate the safety and risks of LLMs, the LLM Safety Leaderboard based on the DecodingTrust framework, which received the NeurIPS 2023 Outstanding Paper Award, has been launched.
This leaderboard evaluates models from the following 8 key perspectives:
- Toxicity and Stereotype Bias
- Adversarial/OOD Robustness
- Privacy
- Machine Ethics and Fairness
Each evaluation category applies specialized Red-teaming algorithms to identify model vulnerabilities, and provides a comprehensive ranking that includes both open and closed models.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Study Finds Safety Training for GPT Models Transforms Rather Than Reduces Sexism
Microsoft AI CEO Criticizes Anthropic: 'Model Welfare' Training Is Dangerous