Pinned post
Using AuditBench? Make sure you have the updated models!
We traced an unexpected hallucination behavior to specific training examples, corrected the data, and retrained the affected models. Here’s what we found and what we checked. 🧵
We traced an unexpected hallucination behavior to specific training examples, corrected the data, and retrained the affected models. Here’s what we found and what we checked. 🧵
New Anthropic Fellows research: Alignment auditing—investigating AI models for unwanted behaviors—is a key challenge for safely deploying frontier models.
We're releasing AuditBench, a suite of 56 LLMs with implanted hidden behaviors to measure progress in alignment auditing.View quoted post on X
