DP61274/IndabaX-Kenya-Tech4MentalHealth-Hackathon

# Tech4MentalHealth — Zindi Basic Needs Basic Rights Kenya · **Tech4MentalHealth**. Multiclass text classification of student statements into `Depression`, `Alcohol`, `Suicide`, `Drugs`. Metric: **multiclass Log Loss**. ## Layout ``` data/ Train.csv, Test.csv, SampleSubmission.csv (download from Zindi) notebooks/ build_notebook.py generator for the baseline notebook 02_Baseline.ipynb reusable pipeline (TF-IDF -> LR -> CV -> OOF -> submission) oof/ out-of-fold probabilities per experiment submissions/ submission CSVs per experiment experiments.csv our own leaderboard ``` ### `experiments.csv` columns `Experiment, TFIDF, Model, Hypothesis, CV_LogLoss, CV_Std, OOF_LogLoss, Submitted, Public_LB, Decision, Notes` - **Hypothesis** — what the run tests and why. - **Submitted** — ❌/✅; which runs consumed a daily Zindi submission. - **Public_LB** — filled in by hand after submitting. - **Decision** — auto against `BEAT_TARGET` (0.62381, the baseline): champion → `Keep (champion)`, beats target → `Candidate`, else `Reject`. Manual values (e.g. `Promote`, `Tune Further`) are preserved. **If a run doesn't beat the target, we don't submit it.** ## Setup ```bash pip install -r requirements.txt ``` Place `Train.csv`, `Test.csv`, `SampleSubmission.csv` in `data/`, then run `notebooks/02_Baseline.ipynb` top-to-bottom (or `jupyter nbconvert --to notebook --execute notebooks/02_Baseline.ipynb`). ## Pipeline ``` Text -> TF-IDF (fit per fold) -> Model -> Stratified 5-Fold CV -> OOF predictions -> Log Loss -> submission ``` Key ideas: - **Stratified 5-Fold, `random_state=42`, frozen `FOLDS`** shared by every experiment so scores are directly comparable. - **OOF predictions** saved for error analysis / calibration / ensembling. - **Class column order is read from `SampleSubmission.csv`** and the model's probability columns are reordered to match — avoids a silently-wrong Log Loss. - No stemming / lemmatization / stopword removal (hurts on short te …