flax/distillation_scripts/run_dataset_sweep.yaml

command:
  - python3
  - ${program}
  - --do_train
  - --do_eval
  - --use_scan
  - --gradient_checkpointing
  - --overwrite_output_dir
  - --predict_with_generate
  - --freeze_encoder
  - --streaming
  - --use_auth_token
  - ${args}
method: grid
metric:
  goal: minimize
  name: gigaspeech-l/validation/wer
parameters:
  model_name_or_path:
    value: distil-whisper/large-32-2
  teacher_model_name_or_path:
    value: openai/whisper-large-v2
  max_train_samples:
    values:
      - 109876
      - 219752
      - 439504
      - 879008
      - 1758015
      - 3516030
      - 7032061
  train_dataset_name:
    value: librispeech_asr-timestamped+librispeech_asr-timestamped+librispeech_asr-timestamped+common_voice_13_0-timestamped+voxpopuli-timestamped+ami-ihm-timestamped+ami-sdm-timestamped+peoples_speech-clean-timestamped+tedlium-timestamped+switchboard-data+gigaspeech-l-timestamped+librispeech_asr-prompted+librispeech_asr-prompted+librispeech_asr-prompted+tedlium-prompted
  train_dataset_config_name:
    value: all+all+all+en+en+ihm+sdm+clean+release3+all+l+all+all+all+release3
  train_split_name:
    value: train.clean.100+train.clean.360+train.other.500+train+train+train+train+train+train+train+train+train.clean.100+train.clean.360+train.other.500+train
  train_dataset_samples:
    value: 2.9+10.4+14.9+89+18.2+10.9+10.9+288+26.8+371.2+226.6+2.9+10.4+14.9+26.8
  eval_dataset_name:
    value: librispeech_asr+librispeech_asr+common_voice_13_0+voxpopuli+ami-ihm+ami-sdm+peoples_speech-clean+tedlium+switchboard-data+gigaspeech-l+spgispeech+chime4+google/fleurs
  eval_dataset_config_name:
    value: all+all+en+en+ihm+sdm+clean+release3+all+l+L+1-channel+en_us
  eval_split_name:
    value: validation.clean+validation.other+validation+validation+validation+validation+validation+validation+validation+validation+validation+validation+validation
  eval_text_column_name:
    value: text+text+text+text+text+text+text+text+text+text+text+text+transcription
  cache_dir:
    value: /home/sanchitgandhi/.cache
  dataset_cache_dir:
    value: /home/sanchitgandhi/.cache
  output_dir:
    value: ./
  per_device_train_batch_size:
    value: 64
  per_device_eval_batch_size:
    value: 64
  dtype:
    value: bfloat16
  learning_rate:
    value: 1e-4
  lr_scheduler_type:
    value: constant_with_warmup
  warmup_steps:
    value: 50
  max_steps:
    value: 10000
  save_steps:
    value: 10001  # don't save checkpoints during sweep
  dataloader_num_workers:
    value: 48
  logging_steps:
    value: 25
  wer_threshold:
    value: 10
program: run_distillation.py
project: distil-whisper-sweeps