CringeBench measures how socially uncalibrated LLM responses are — sycophancy, forced humour, purple prose, robotic disclaimers, and general second-hand embarrassment.
Every model is asked the same set of prompts designed to surface performative or self-aggrandizing behaviour. Every response is then scored by every model acting as a judge, producing an N×N cross-evaluation matrix.
for each model M:
for… See the full description on the dataset page:
https://huggingface.co/datasets/av-codes/cringebench.
