wassimhaoues/TUN_STT_MODEL

This is a speech-to-text model fine-tuned for Tunisian Derja using 30 hours of speech data, built on the Whisper Small base model. # TUN STT Model Development workspace for preparing and fine-tuning Whisper Small to better transcribe Tunisian Derja. ## Quick Start ```bash python3 -m venv .venv source .venv/bin/activate python -m pip install --upgrade pip python -m pip install -r requirements/runtime.txt -r requirements/dev.txt pre-commit install ``` ## Common Commands ```bash make lint make format make test make validate-dataset make check-dataset make smoke-whisper make baseline make train-smoke make eval-checkpoint ``` ## Baseline Tracking The baseline script now records a tracked experiment report in `reports/` every time it runs. ```bash python training/baseline_test.py python training/baseline_test.py --run-name baseline-start --notes "raw whisper-small before fine-tuning" ``` Each run writes: - `reports/experiment_history.csv`: one row per experiment run - `reports/runs/ /summary.md`: readable baseline summary - `reports/runs/ /predictions.csv`: per-sample predictions for that run ## Phase 02 Smoke Training Phase 02 adds a reproducible smoke fine-tuning pipeline for `openai/whisper-small`. It trains on a deterministic subset, saves checkpoints under `outputs/`, and writes tracked run metadata under `reports/`. ```bash python training/train_whisper_small.py python training/train_whisper_small.py --train-samples 256 --valid-samples 64 --max-steps 20 python training/train_whisper_small.py --run-name whisper-small-phase02-gpu --notes "first laptop smoke run" ``` Default smoke behavior: - train subset: `1000` rows - valid subset: `200` rows - precision: `auto` with GPU-aware fallback - max steps: `60` - batch config: `4 x grad_accum 4` Each smoke run writes: - `outputs/train_runs/ /`: checkpoints, trainer state, logs - `reports/runs/ /summary.md`: tracked Phase 02 report - `reports/runs/ /training_config.json`: exact config snapshot - `reports/runs/ /environment.json`: hardware and precision snapshot - `reports/runs/ /selected_train_manifest.csv`: exact train subset used - `re …