config.yaml

hydra:
  output_subdir: null  
  run:
    dir: .
  sweep:
    dir: .
    subdir: .
  job_logging:
    root:
      level: INFO
  job:
    env_set:
      TOKENIZERS_PARALLELISM: "false"


defaults:
  - base_config  # see src/arguments.py
  - _self_
  - override hydra/hydra_logging: disabled  
  - override hydra/job_logging: disabled  


wandb:
  log: true
  entity: ${user}  # Change this to your wandb username.
  project: ccm
  group: ${basename:${data.dataset_name}}/${basename:${model.model_name_or_path}}-${training.comp.comp_type}-ntok${training.comp.num_comp_tokens}
  name: ${basename:${model.model_name_or_path}}-${wandb.tag}

model:
  pretrained: true

training:
  predict_with_generate: true
  generation_max_length: 128

  do_train: true
  do_eval: true

  report_to: "none"  # THIS MUST BE NONE. Use wandb args to control logging.

  dataloader_num_workers: 0  # If > 0, some weird process hanging might occur.

  # Default training params: effective batch size = 16
  num_train_epochs: 3
  fp16: false
  fp16_full_eval: false
  per_device_train_batch_size: 8
  per_device_eval_batch_size: 8
  gradient_accumulation_steps: 2

  # Save/eval every 1000 steps and track best model
  overwrite_output_dir: false  # Resume training from checkpoint if it exists.
  evaluation_strategy: steps
  save_strategy: steps
  eval_steps: 1000
  save_steps: 1000
  save_total_limit: 1
  load_best_model_at_end: false # Always save the best model, eval_steps should be equal to save_steps
  logging_steps: 200

  # metric_for_best_model: unseen_rougeL
  greater_is_better: true
  max_source_length: 256
  max_target_length: 256