command:
  - python3
  - ${program}
  - --do_train
  - --do_eval
  - --use_scan
  - --gradient_checkpointing
  - --overwrite_output_dir
  - --predict_with_generate
  - ${args}
method: random
metric:
  goal: minimize
  name: eval/wer
parameters:
  model_name_or_path:
    value: distil-whisper/large-32-2
  dataset_name:
    value: distil-whisper/librispeech_asr
  dataset_config_name:
    value: all
  train_split_name:
    value: train.clean.100+train.clean.360+train.other.500
  eval_split_name:
    value: validation.clean
  text_column_name:
    value: whisper_transcript
  cache_dir:
    value: /home/sanchitgandhi/cache
  dataset_cache_dir:
    value: /home/sanchitgandhi/cache
  output_dir:
    value: ./
  per_device_train_batch_size:
    value: 32
  per_device_eval_batch_size:
    value: 16
  dtype:
    value: bfloat16
  learning_rate:
    distribution: log_uniform
    max: -6.91
    min: -11.51
  warmup_steps:
    value 500
  num_train_epochs:
    value: 1
  preprocessing_num_workers:
    value: 16
  dataloader_num_workers:
    value: 16
  logging_steps:
    value: 25
  freeze_encoder:
    values:
      - True
      - False

program: run_finetuning.py
project: distil-whisper