flax/long_form_transcription_scripts/run_length_penalty_sweep.yaml

command:
  - python3
  - ${program}
  - --streaming
  - ${args}
method: grid
metric:
  goal: minimize
  name: tedlium-long-form/validation/wer
parameters:
  model_name_or_path:
    value: sanchit-gandhi/large-32-2-ts-freeze-28k-wer-10
  subfolder:
    value: checkpoint-15000
  dataset_name: 
    value: distil-whisper/tedlium-long-form
  dataset_config_name: 
    value: all
  dataset_split_name: 
    value: validation
  cache_dir: 
    value: /home/sanchitgandhi/.cache
  dataset_cache_dir: 
    value: /home/sanchitgandhi/.cache
  output_dir: 
    value: ./
  wandb_dir: 
    value: /home/sanchitgandhi/.cache
  per_device_eval_batch_size:
    value: 32
  dtype: 
    value: bfloat16
  report_to:
    value: wandb
  generation_num_beams:
    value: 5
  generation_max_length:
    value: 256
  length_penalty:
    values:
      - 0.6
      - 0.8
      - 1.0
      - 1.2
      - 1.4
program: run_long_form_transcription.py
project: distil-whisper-long-form