EXPRESS: 33,679 Reddit Experiences with Self-Disclosed Emotion Labels

33,679 naturally occurring Reddit-based human experiences paired with self-disclosed emotion labels form the EXPRESS benchmark. Created by author bangzhao and detailed in a 2025 arXiv paper, this dataset uses emotions explicitly disclosed by the original authors as ground-truth labels. The dataset was last updated on Hugging Face in May 2026.

Use Cases

• Benchmarking language model performance on fine-grained emotion recognition based on self-disclosed labels.
• Training emotion classifiers on naturally occurring human experiences sourced from Reddit.
• Analyzing the relationship between narrative content and self-reported emotional states.
• Studying the reliability of author-disclosed emotions as ground-truth labels for machine learning.

Strengths

• Contains 33,679 data points, providing a substantial corpus for training and evaluation.
• Uses author-disclosed emotions as ground-truth labels, which may offer high validity for the stated emotional context.
• Based on naturally occurring human experiences from Reddit, likely reflecting diverse and authentic language.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific data formats, file sizes, and licensing information are unavailable.
• Data may reflect the temporal, demographic, and linguistic biases inherent to its Reddit source.

Provenance

Source

Reddit

Collection Method

Likely gathered via web scraping or API collection of user posts.

Freshness

Last updated 2026-05-03 05:00:54; freshness should be verified.