lora_config:
  bias: none
  lora_alpha: 64
  lora_dropout: 0.05
  r: 32
  target_modules:
  - q_proj
  - v_proj
num_workers: 4
pretrained_model: openai/whisper-large-v3
train:
  huggingface_load:
  - name: audio_1h
    path: evie-8/kinyarwanda-speech-hackathon
    split: train
training_args:
  eval_steps: 200
  eval_strategy: steps
  fp16: true
  generation_max_length: 200
  gradient_accumulation_steps: 2
  gradient_checkpointing: true
  gradient_checkpointing_kwargs:
    use_reentrant: false
  greater_is_better: false
  hub_model_id: patrickcmd/whisper-large-v3-kin-1h
  learning_rate: 1.0e-05
  load_best_model_at_end: true
  logging_steps: 200
  max_steps: 10000
  metric_for_best_model: loss
  per_device_eval_batch_size: 8
  per_device_train_batch_size: 8
  predict_with_generate: true
  push_to_hub: true
  save_steps: 1000
  save_total_limit: 2
  warmup_steps: 100
use_peft: false
validation:
  huggingface_load:
  - path: jq/kinyarwanda-speech-hackathon
    split: dev_test[:300]
  source:
    language:
    - kin
    preprocessing:
    - set_sample_rate:
        rate: 16000
    type: speech
  target:
    language:
    - kin
    preprocessing:
    - lower_case
    type: text