MontaCHOUIG/whisper-tunisian-ft

# Tunisian Arabic Whisper Large-v3 LoRA Fine-tuning Production-ready fine-tuning pipeline for Tunisian Arabic ASR with OpenAI Whisper Large-v3, Hugging Face Transformers, Datasets, Evaluate, and PEFT LoRA. ## Installation ```bash python -m venv venv venv\Scripts\activate pip install -r requirements.txt ``` On Linux/macOS, activate with `source venv/bin/activate`. Install a CUDA-compatible PyTorch build if your environment needs a specific CUDA version. ## Dataset Structure Training expects parquet shards under the path configured in `configs/train.yaml`, defaulting to: ```text /data/augmented_data/ data-00000.parquet data-00001.parquet ... ``` Nested parquet files are also discovered recursively. Each row must contain: ```text audio_id: string audio: { bytes: binary WAV, sampling_rate: int64 } transcript: string segments: list[...] # ignored during training source: string duration: float ``` The `transcript` column is used as the target text. The `segments` column is removed during preprocessing. ## Configuration All training, dataset, model, LoRA, generation, and output settings live in `configs/train.yaml`. Important fields: - `dataset.path`: parquet directory, default `/data/augmented_data` - `dataset.train_test_split`: test ratio, default `0.1` - `dataset.seed`: reproducible shuffle and split seed - `training.fp16` / `training.bf16`: choose one mixed precision mode - `lora.target_modules`: default `q_proj` and `v_proj` - `output.save_merged_model`: optionally save a merged full model ## Training ```bash python src/train.py --config configs/train.yaml ``` The script logs dataset size, train/test size, average and maximum duration, GPU information, LoRA parameter counts, final metrics, and total training time. Outputs are written under `outputs/whisper-large-v3-tunisian-arabic-lora` by default: ```text adapter_config.json adapter_model.safetensors tokenizer files preprocessor_config.json generation_config.json training_args.json metrics.js …