Transluce Releases WeirdChat: A Catalog of 175K Strange LLM Behaviors

ChowdhuryNeil · x · 2026-07-31

Transluce has released WeirdChat, a public catalog of over 175,000 annotated transcripts designed to systematically study unexpected or harmful behaviors in frontier large language models.

Using automated elicitation tools, the project uncovered over 1,300 behavioral patterns across multiple open-weight models including DeepSeek-V4-Flash, Gemma 4 31B, Nemotron 3 Ultra, and Qwen3.6. These behaviors range from user harm and inappropriate actions to misrepresentation of actions and harmful advice. The dataset aims to help developers and researchers better understand and anticipate model failures in real-world deployments.

Related event: Transluce Launches WeirdChat with 175k Annotated AI Anomalies(2 posts)→

Original post →

More from Models

Models channel →