callbacks:
  callbacks: []
data:
  datamodule:
    _target_: classy.data.data_modules.ClassyDataModule
    task: ${task}
    dataset_path: data/aida
    train_dataset:
      _target_: classy.data.dataset.hf.classification.HFQADataset.from_file
      transformer_model: ${transformer_model}
      additional_special_tokens: ${model.additional_special_tokens}
      min_length: 5
      max_length: 500
      tokens_per_batch: 2000
      max_batch_size: 10
      section_size: 10000
      prebatch: true
      materialize: false
      for_inference: false
    validation_dataset:
      _target_: classy.data.dataset.hf.classification.HFQADataset.from_file
      transformer_model: ${transformer_model}
      additional_special_tokens: ${model.additional_special_tokens}
      min_length: 5
      max_length: 500
      tokens_per_batch: 2000
      max_batch_size: 10
      section_size: 10000
      prebatch: true
      materialize: true
      for_inference: true
    validation_split_size: 0.1
    test_split_size: 0.1
    max_nontrain_split_size: 10000
    shuffle_dataset: true
device:
  gpus:
  - 0
  precision: 32
  amp_level: O0
model:
  _target_: classy.pl_modules.hf.classification.HFQAPLModule
  transformer_model: ${transformer_model}
  additional_special_tokens: []
  optim_conf:
    _target_: classy.optim.factories.RAdamFactory
    lr: 1.0e-05
    weight_decay: 0.01
    no_decay_params:
    - bias
    - LayerNorm.weight
prediction:
  dataset:
    _target_: classy.data.dataset.hf.classification.HFQADataset.from_samples
    transformer_model: ${transformer_model}
    additional_special_tokens: ${model.additional_special_tokens}
    min_length: -1
    max_length: -1
    tokens_per_batch: 800
    max_batch_size: -1
    section_size: 10000
    prebatch: true
    materialize: false
    for_inference: true
training:
  seed: 12
  pl_trainer:
    _target_: pytorch_lightning.Trainer
    accumulate_grad_batches: 4
    gradient_clip_val: 10.0
    val_check_interval: 1.0
    max_steps: 1000000
  early_stopping_callback:
    _target_: pytorch_lightning.callbacks.EarlyStopping
    monitor: ${callbacks_monitor}
    mode: ${callbacks_mode}
    patience: 25
  model_checkpoint_callback:
    _target_: classy.pl_callbacks.best_checkpoint.ModelCheckpointWithBest
    monitor: ${callbacks_monitor}
    mode: ${callbacks_mode}
    verbose: true
    save_top_k: 3
    dirpath: checkpoints
    save_last: true
  resume_from: null
logging:
  wandb:
    use_wandb: true
    project_name: esc-ed
    experiment_name: aida-longformer-large-*sep-gam-cand-shuffle
    anonymous: null
    run_id: null
task: qa
project_name: classy
exp_name: esc-aida-longformer-large-gam-cand-shuffle
exp_folder: ./experiments/${exp_name}
transformer_model: bert-base-cased
callbacks_monitor: val_accuracy
callbacks_mode: max
profiles:
  supported_tasks:
  - qa
  - sentence-pair
  - sequence
  - token
  - generation
  transformer_model: allenai/longformer-large-4096
  candidates_separator: '*'
  training:
    pl_trainer:
      accumulate_grad_batches: 8
      val_check_interval: 2048
      max_steps: 100000
  model:
    _target_: src.esc_ed_module.ESCModule
    additional_special_tokens: []
    transformer_model: ${transformer_model}
    attention_window: 64
    modify_global_attention: true
    optim_conf:
      _target_: classy.optim.factories.RAdamFactory
      lr: 1.0e-05
      weight_decay: 0.01
      no_decay_params:
      - bias
      - LayerNorm.weight
  data:
    datamodule:
      train_dataset:
        _target_: src.data.esc_ed_dataset.ESCEDDataset.from_file
        transformer_model: ${transformer_model}
        additional_special_tokens: ${model.additional_special_tokens}
        candidates_separator: ${candidates_separator}
        shuffle_candidates_prob: 0.0
        min_length: 0
        max_length: 1024
        tokens_per_batch: 1024
        max_batch_size: 10
        section_size: 20000
        prebatch: true
        materialize: false
        for_inference: false
      validation_dataset:
        _target_: src.data.esc_ed_dataset.ESCEDDataset.from_file
        transformer_model: ${transformer_model}
        additional_special_tokens: ${model.additional_special_tokens}
        candidates_separator: ${candidates_separator}
        min_length: 0
        max_length: 1024
        tokens_per_batch: 2048
        max_batch_size: 10
        section_size: 10000
        prebatch: true
        materialize: true
        for_inference: false
      shuffle_dataset: true
  prediction:
    dataset:
      _target_: src.data.esc_ed_dataset.ESCEDDataset.from_samples
      transformer_model: ${transformer_model}
      additional_special_tokens: ${model.additional_special_tokens}
      candidates_separator: ${candidates_separator}
      min_length: -1
      max_length: -1
      tokens_per_batch: 2048
      max_batch_size: -1
      section_size: 10000
      prebatch: true
      materialize: false
      for_inference: true