Peter-Abugri/tech4mentalhealth

4th-place solution to the Zindi/IndabaX Kenya Tech4MentalHealth NLP challenge — RoBERTa ensemble with grouped cross-validation. # Tech4MentalHealth — NLP Classification A four-class text classifier that categorises statements written by Kenyan university students into **Depression, Alcohol, Suicide, or Drugs**, built for the Zindi / IndabaX Kenya *Tech4MentalHealth* challenge. The eventual goal is a mental-health chatbot prototype that routes a student's message to the right kind of support. **Result:** 4th place, log loss **0.359**. ## Approach - **Baseline:** TF-IDF (word + character n-grams) with logistic regression. - **Transformers:** fine-tuned RoBERTa-base and RoBERTa-large, each averaged over three random seeds to reduce variance on the small dataset. - **Ensembling:** geometric blend of the models, with weights chosen on out-of-fold predictions. - **Validation:** 5-fold stratified *group* cross-validation (grouped on normalised text to prevent duplicate leakage). Local CV tracked the leaderboard to within ~0.02 throughout. See `documentation.md` for a full, beginner-friendly write-up of the data, the metric, and every modelling decision. ## Data The dataset is the property of Zindi and is **not included** in this repository, per the competition rules. Download it from the competition page if you are a participant, and place the CSV files in a local `data/` folder (git-ignored). ## Repository ``` solution.ipynb model training, blending, and EDA (Colab notebook) documentation.md full project write-up README.md ``` ## Tech Python · scikit-learn · PyTorch · Hugging Face Transformers · pandas