MentalDistress: A multi-class social media text dataset for mental health–related emotion detection

MentalDistress is a human preference dataset developed for evaluating personalized alignment in Natural Language Processing(NLP). This dataset is developed to support mental health text classification research in the English language. It consists of manually curated and annotated English text samples categorized into five psychological states. The dataset is designed to facilitate supervised learning approaches for detecting emotional distress and high-risk mental health indicators from textual data.

The corpus contains a total of 10,100 annotated text samples, distributed across five classes representing different mental health conditions.

# Class Distribution
The dataset includes the following categories:
- Suicidal: 2,171 samples
- Depressed: 2,051 samples
- Anxious: 2,038 samples
- Frustrated: 2,062 samples
- Others: 1,778 samples
The class distribution is relatively balanced, making the dataset suitable for multi-class classification experiments without severe class imbalance issues.

# Data Collection and Annotation
The text samples were collected from publicly available sources and manually reviewed. Each instance was carefully annotated according to predefined psychological category guidelines to ensure labeling consistency. Quality control measures were applied to maintain annotation reliability.
# Key Features
- Five-class mental health categorization
- Manually annotated dataset
- Fairly balanced class distribution
- Suitable for classical ML and deep learning models

# Potential Use Cases
- Mental health text classification
- Early detection of psychological distress
- NLP research
- Transformer-based model benchmarking
- AI-assisted mental health screening research

# File Format
The dataset is provided in CSV format with the following columns:
- Text: English textual content
- Label: Class name (Anxious, Depressed, Frustrated, Others, Suicidal)