Biomedical Text Dataset for Early Autism Prediction Using Transformer Models

This dataset contains curated and augmented textual data related to Autism Spectrum Disorder (ASD), compiled from publicly available sources including the Mendeley repository, caregiver narratives, and clinical discussion forums. The data was filtered to remove duplicates and retain only ASD-relevant content, preserving the natural linguistic patterns present in parental language. Synthetic data generation using TinyLLaMA was employed to balance class distribution and expand coverage of DSM-5-aligned symptom categories. The final dataset supports both binary classification (ASD vs. Non-ASD) with 5,000 samples and multi-class classification with 500 ASD-positive samples across 10 symptom categories. This resource is intended for use in domain-specific natural language processing, mental health diagnostics, and ASD symptom modeling research.