run_sweep.yaml

command:
  - python3
  - ${program}
  - --load_best_model_at_end
  - --fp16
  - --fp16_full_eval
  - --do_train
  - --do_eval
  - --trust_remote_code
  - --overwrite_output_dir
  - --ignore_mismatched_sizes
  - --gradient_checkpointing
  - ${args}
method: grid
metric:
  goal: maximize
  name: eval/accuracy
parameters:
  model_name_or_path:
    values:
      - facebook/mms-lid-126
      - openai/whisper-large-v3
  train_dataset_name:
    value: sanchit-gandhi/vctk+facebook/voxpopuli+sanchit-gandhi/edacc
  train_dataset_config_name:
    value: default+en_accented+default
  train_split_name:
    value: train+test+validation
  train_label_column_name:
    value: accent+accent+accent
  eval_dataset_name:
    value: sanchit-gandhi/edacc
  eval_dataset_config_name:
    value: default
  eval_split_name:
    value: test
  eval_label_column_name:
    value: accent
  output_dir:
    value: ./
  remove_unused_columns:
    value: false
  learning_rate:
    value: 1e-4
  lr_scheduler_type:
    value: constant_with_warmup
  max_length_seconds:
    value: 10  # give some data diversity for longer audio samples
  min_length_seconds:
    value: 5
  attention_mask:
    value: false
  warmup_steps:
    value: 50
  max_steps:
    value: 2000
  per_device_train_batch_size:
    value: 32
  per_device_eval_batch_size:
    value: 16
  preprocessing_num_workers:
    value: 16
  dataloader_num_workers:
    value: 4
  logging_strategy:
    value: steps
  logging_steps:
    value: 10
  evaluation_strategy:
    value: steps
  eval_steps:
    value: 2000
  save_strategy:
    value: steps
  save_steps:
    value: 2000
  metric_for_best_model:
    value: accuracy
  freeze_base_model:
    value: false
  push_to_hub:
    value: false
program: run_audio_classification.py
project: mms-lid-accent-classification