This dataset accompanies the paper Detecting High-Stakes Interactions with Activation Probes, presented at the ICML 2025 Workshop on Actionable Interpretability, accepted to NeurIPS 2025.
Every sample is a user-facing LLM interaction labelled as high-stakes or low-stakes. The label reflects whether the conversation involves potentially consequential outcomes (medical advice, legal matters, financial decisions, etc.) vs. routine queries.
The… See the full description on the dataset page:
https://huggingface.co/datasets/Arrrlex/models-under-pressure.
