configs/base_config.yaml

model:
  id: "openai/whisper-large-v3-turbo"
  task: "transcribe"
  max_new_tokens: 128
  chunk_length_s: 30

training:
  output_dir: "./adapters"
  per_device_train_batch_size: 4
  gradient_accumulation_steps: 4
  warmup_steps: 200
  max_steps: 4000
  save_steps: 500
  eval_steps: 500
  learning_rate: 1.0e-4
  fp16: true
  # CRITICAL on Windows: multiprocessing spawn breaks with tokenizers
  dataloader_num_workers: 0

audio:
  sample_rate: 16000
  max_duration_s: 30
  noise_snr_db_range: [5, 20]
  augmentation_prob: 0.6

paths:
  data_cache: "./data_cache"
  adapters: "./adapters"
  models: "./models"
  noise_samples: "./noise_samples"