flax/distillation_scripts/run_mse_sweep.yaml

command:
  - python3
  - ${program}
  - --do_train
  - --do_eval
  - --gradient_checkpointing
  - --overwrite_output_dir
  - --predict_with_generate
  - --streaming
  - --use_auth_token
  - --use_scan
  - ${args}
method: grid
metric:
  goal: minimize
  name: eval/wer
parameters:
  model_name_or_path:
    value: distil-whisper/large-16-2
  teacher_model_name_or_path:
    value: openai/whisper-large-v2
  train_dataset_name:
    value: librispeech_asr+librispeech_asr+librispeech_asr+common_voice_13_0+voxpopuli+ami-ihm+ami-sdm+peoples_speech-clean+tedlium+switchboard-data+gigaspeech-l+spgispeech
  train_dataset_config_name:
    value: all+all+all+en+en+ihm+sdm+clean+release3+all+l+L
  train_split_name:
    value: train.clean.100+train.clean.360+train.other.500+train+train+train+train+train+train+train+train+train
  train_dataset_samples:
    value: 100+360+500+2300+450+90+90+12000+450+3600+2500+5000
  eval_dataset_name:
    value: "distil-whisper/gigaspeech-l"
  eval_dataset_config_name:
    value: "l"
  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: 64
  dtype:
    value: bfloat16
  learning_rate:
    value: 0.0001
  lr_scheduler_type:
    value: constant_with_warmup
  warmup_steps:
    value: 50
  max_steps:
    value: 2500
  eval_steps:
    value: 2500
  save_steps:
    value: 2001  # don't save checkpoints during sweep
  dataloader_num_workers:
    value: 16
  logging_steps:
    value: 5
  wer_threshold:
    value: 10
  mse_weight:
    values:
      - 0.0
      - 0.3
      - 1
      - 3
program: run_distillation.py
project: distil-whisper-sweeps