Fine-tuned transformer models (DistilBERT, BERT, RoBERTa) on a labeled 6-class Reddit dataset (~6,000 balanced samples) to detect stress, anxiety, depression, and related conditions.

Achieved an F1 score of 0.73 and accuracy of 74.8% using RoBERTa, optimized via grid search over learning rate, batch size, and epochs.

Applied LIME for interpretability and explored additional TF-IDF-based models for baseline comparison.

Tools: Python, Hugging Face Transformers, PyTorch, Scikit-learn, LIME, Matplotlib, Google Colab