Emotion-probes is a synthetic text dataset designed for research on emotion understanding and model interpretability. It aims to extract emotion vectors and emotion masking capabilities from models and is widely used in scenarios such as emotion classification, model alignment, security research, and analysis of the internal mechanisms of large models. The dataset contains approximately 447,000 samples, each containing fields such as actual emotion, expressed emotion, text content, and role information.

Data composition

The data consists of four parts:

• expression/stories.parquet: 205,200 short stories covering 171 emotions and 100 themes.
• deflection/dialogues.parquet: Approximately 239,400 dialogues that mask emotions.
• expression/neutral_stories.parquet: 1200 sentiment-neutral stories used to eliminate confounding factors in principal component analysis.
• deflection/neutral_dialogues.parquet: 1200 neutral/AI dialogues for eliminating confounding factors in principal component analysis.