dump_targets_config.yaml

#file: noinspection LongLine
defaults:
  - generation@generation_config: generation_config
  - generation/logits_processor@logits_processor_list: ???
  - generation/stopping_criteria@stopping_criteria_list: ???
  - tokenizer: autotokenizer
  - collator: dump_targets
  - dataset: ???
  - _self_

seed: 1234
pretrained_model_name_or_path: ???
pretrained_model_revision: null
output_path: ???  # Where to save the distillation dataset
save_interval: 100  # Save dataset every N samples (for incremental saving)

eos_token_id: ${get_tokenizer_eos_token_id:${tokenizer}}
# Generation parameters
max_length: null
max_new_tokens: null
batch_size: 1  # Batch size for generation

# Dataloader configuration
dataloader:
  _target_: torch.utils.data.DataLoader
  batch_size: ${batch_size}
  num_workers: 0  # Set to 0 for inference to avoid multiprocessing issues
  pin_memory: true
  shuffle: false  # No shuffle for inference
  drop_last: false  # Don't drop last batch

backend: ${set_backend:}
gen_kwargs:
  generation_config: ${generation_config}
  logits_processor:
    _target_: src.custom_transformers.generation.HydraCompatibleLogitsProcessorList
    logits_processor_dict: ${logits_processor_list}
  stopping_criteria:
    _target_: src.custom_transformers.generation.HydraCompatibleStoppingCriteriaList
    stopping_criteria_dict: ${stopping_criteria_list}
  max_length: ${max_length}
  max_new_tokens: ${max_new_tokens}