Mental Health Text Classification Dataset (4-Class)

Dataset Description

This dataset contains short, user‑generated texts labeled for 4‑class mental health classification: Suicidal, Depression, Anxiety, and Normal. It is a derived dataset created by combining and cleaning three public mental‑health corpora, then re‑labeling them into a unified 4‑class scheme and exporting CSV files suitable for both classical ML and modern NLP models.

The repository includes:

An unbalanced main training corpus (realistic class skew).
A strictly balanced test split for fair evaluation.
A feature‑engineered file with basic text statistics (length, URLs, emojis, punctuation, etc.).

Important: This dataset is intended for research and education only. It is not a clinical tool and must not be used for real‑world diagnosis, triage, or crisis intervention.

Dataset Structure

Files

mental_heath_unbanlanced.csv
- Main training corpus with 48,945 samples and a realistic, unbalanced class distribution (Normal, Depression, Suicidal, Anxiety).
- Columns:
  - text – Cleaned post or statement text.
  - status – Final label in {Suicidal, Depression, Anxiety, Normal}.
mental_health_combined_test.csv
- Balanced test split with 992 samples (exactly 248 per class).
- Columns:
  - text – Cleaned post.
  - status – Final 4‑class label.
- Built by combining held‑out data from the sources, applying the same cleaning and label mapping, removing duplicates, and downsampling each class to equal size for fair evaluation.
mental_health_feature_engineered.csv
- Cleaned and feature‑engineered subset for classical models and analysis.
- Columns:
  - Core: Unique_ID, text, status (4‑class label).
  - Features:
    - text_length – Number of characters.
    - word_count – Number of whitespace‑separated tokens.
    - num_urls – Count of URL patterns.
    - num_emojis – Count of emoji characters.
    - num_special_chars – Count of non‑alphanumeric characters.
    - num_excess_punct – Count of repeated punctuation sequences (e.g., !!!, ???, ...).
    - avg_word_length – Average characters per word.

Splits

This repo is organized at the file level (no HF train/test config baked in). A common convention is:

mental_heath_unbanlanced.csv → train (and derive your own validation split).
mental_health_combined_test.csv → test.

Example with datasets:

from datasets import load_dataset

ds_train = load_dataset( "your-username/Mental-Health_Text-Classification_Dataset", data_files={"train": "mental_heath_unbanlanced.csv"}, )

ds_test = load_dataset( "your-username/Mental-Health_Text-Classification_Dataset", data_files={"test": "mental_health_combined_test.csv"}, )

print(ds_train) print(ds_test)

text

For the feature file:

ds_fe = load_dataset( "your-username/Mental-Health_Text-Classification_Dataset", data_files={"feature": "mental_health_feature_engineered.csv"}, )

text

Source Data and Provenance

This dataset does not collect new data from individuals. Instead, it is built by downloading, cleaning, and merging three existing public resources.[file:64][file:65]

Source datasets:

Suicide and Depression Detection (Kaggle) – Nikhileswar Komati
Reddit posts from suicide‑related communities labeled as suicidal vs non‑suicidal / depression vs control.
https://www.kaggle.com/datasets/nikhileswarkomati/suicide-watch
Sentiment Analysis for Mental Health (Kaggle) – Suchintika Sarkar
Short statements labeled as normal, depression, suicide, anxiety, stress, bipolarity, personality disorder, etc.
https://www.kaggle.com/datasets/suchintikasarkar/sentiment-analysis-for-mental-health
Reddit Mental Health Classification (Murarka et al.)
Dataset released with “Detection and Classification of Mental Illnesses on Social Media using RoBERTa.”
Reddit posts labeled into ADHD, Anxiety, Bipolar, Depression, PTSD, and None.
https://github.com/amurark/mental-health-detection

Only the processed, relabeled, and re‑split samples are redistributed here. Users must obtain the original raw data from the links above if needed.

Preprocessing and Label Mapping

Cleaning

All source datasets are passed through a common cleaning pipeline:

Drop rows with missing text or labels (status).
Deduplicate texts across and within sources.
Apply regex‑based text normalization (whitespace cleanup, formatting fixes).
Filter out extremely short and excessively long texts to remove non‑informative samples.

Unified 4‑Class Labels

Original labels vary widely (suicide vs non, multiple disorders, control). They are mapped to a four‑class scheme:

Suicidal – Posts explicitly labeled as suicide / suicidal ideation.
Depression – Posts labeled as depression / depressed.
Anxiety – Posts labeled with anxiety‑related categories.
Normal – Control / non‑suicide / normal posts.

Samples whose labels cannot be clearly mapped (e.g., stress, bipolar, personality disorder, PTSD, mixed or ambiguous labels) are discarded to keep the final label space clean.

Splits and Balance

The main training file (mental_heath_unbanlanced.csv) preserves the natural class imbalance after cleaning and mapping.
The balanced test file (mental_health_combined_test.csv) is created by:
- Building a held‑out pool from the sources.
- Applying the same cleaning and label mapping.
- Removing duplicates.
- Downsampling each class to 248 samples for a total of 992 rows.

The feature‑engineered file is generated from the cleaned sentiment subset by adding the numerical text features listed above.

Intended Uses

Primary Uses

Research and teaching on mental‑health‑related text classification.
Benchmarking binary or multi‑class classifiers (especially 4‑class).
Studying the impact of class imbalance, basic feature engineering, and model robustness on noisy social‑media text.

Typical tasks:

4‑way text classification: Suicidal vs Depression vs Anxiety vs Normal.
Ablation studies: raw text vs text + simple features (from the feature‑engineered file).
Transfer learning experiments on mental‑health Reddit data.

Out‑of‑Scope Uses

Clinical decision‑making, diagnosis, or risk assessment.
Any deployment that might influence real‑world medical, therapeutic, or crisis‑intervention workflows.
Individual‑level profiling, surveillance, or moderation without appropriate human review and safeguards.

Ethical Considerations and Limitations

Non‑clinical labels: Labels are derived from subreddit context and dataset creators’ heuristics, not formal diagnoses by clinicians.
Domain bias: Content is mostly from Reddit and similar platforms; models may not generalize to clinical notes, private messages, or other domains.
Demographic unknowns: There is no reliable demographic metadata; potential demographic, cultural, or linguistic biases cannot be quantified.
No guarantee of harm‑free samples: Text may contain distressing content, including mentions of self‑harm or suicide; users should handle and share it with care, especially in teaching settings.

Users are encouraged to:

Use this dataset only in controlled research/educational environments.
Avoid building systems that auto‑label individual users in high‑stakes contexts without expert oversight.
Follow your institution’s ethics and IRB/IEC guidelines where applicable.

Citation

Please cite the original datasets and this derived dataset.

Original datasets

Komati, N. Suicide and Depression Detection [Dataset]. Kaggle.
https://www.kaggle.com/datasets/nikhileswarkomati/suicide-watch
Sarkar, S. Sentiment Analysis for Mental Health [Dataset]. Kaggle.
https://www.kaggle.com/datasets/suchintikasarkar/sentiment-analysis-for-mental-health
Murarka, A., Radhakrishnan, B., & Ravichandran, S. (2021). Detection and Classification of Mental Illnesses on Social Media using RoBERTa [Dataset and code]. GitHub.
https://github.com/amurark/mental-health-detection

This derived dataset

Mukherjee, P. (2025). Mental Health Text Classification Dataset (4‑Class) [Dataset]. Hugging Face Hub. https://huggingface.co/datasets/ourafla/Mental-Health_Text-Classification_Dataset.