mode: unlearn
task_name: custom_hf_code_unit_qwen2_5_coder_3b_prod
model:
  model_args:
    pretrained_model_name_or_path: Qwen/Qwen2.5-Coder-3B
    attn_implementation: flash_attention_2
    torch_dtype: bfloat16
    peft_name: dbaysal/qwen2.5coder-3b-learned
    peft_checkpoint_subfolder: checkpoint-282
  tokenizer_args:
    pretrained_model_name_or_path: Qwen/Qwen2.5-Coder-3B
  template_args:
    apply_chat_template: false
    system_prompt: You are a helpful assistant.
    system_prompt_with_special_tokens: '<|begin_of_text|><|start_header_id|>system<|end_header_id|>


      You are a helpful assistant.<|eot_id|>'
    user_start_tag: ''
    user_end_tag: ''
    asst_start_tag: ''
    asst_end_tag: ''
    date_string: 10 Apr 2025
trainer:
  handler: PROD
  args:
    per_device_train_batch_size: 4
    per_device_eval_batch_size: 16
    gradient_accumulation_steps: 8
    learning_rate: 0.0003
    bf16: true
    bf16_full_eval: false
    logging_steps: 1
    output_dir: ./saves/unlearn/custom_hf_code_unit_qwen2_5_coder_3b_prod
    logging_dir: ./saves/unlearn/custom_hf_code_unit_qwen2_5_coder_3b_prod/logs
    report_to: tensorboard
    ddp_find_unused_parameters: None
    gradient_checkpointing: false
    optim: adamw_torch
    save_strategy: epoch
    save_only_model: true
    weight_decay: 0.0
    do_train: true
    do_eval: false
    eval_on_start: false
    eval_strategy: 'no'
    num_train_epochs: 3
    seed: 42
    save_total_limit: null
    overwrite_output_dir: true
    lr_scheduler_type: constant
    fp16: false
    remove_unused_columns: false
  method_args:
    gamma: 1.0
    alpha: 0.0
    retain_loss_type: NLL
    top_p: 0.8
    temperature: null
    'N': 1
    max_N: null
    prod_alpha: 0.0
    reference_model_args:
      load_in_8bit: true
data:
  forget:
    CustomHF_forget:
      handler: CompletionDataset
      args:
        hf_args:
          path: dbaysal/forget
          split: train
        text_key: code
        max_length: 2056
        prefix_key: check
  retain:
    CustomHF_retain:
      handler: CompletionDataset
      args:
        hf_args:
          path: dbaysal/retain-half
          split: train
        text_key: code
        max_length: 2056
        prefix_key: check
  anchor: forget
  batch_mode: paired
  batch_order: random
collator:
  DataCollatorForSupervisedDataset:
    handler: DataCollatorForSupervisedDataset
    args:
      padding_side: left
codecarbon:
  enabled: true
  project_name: custom_hf_code_unit_qwen2_5_coder_3b_prod
  output_dir: ./saves/unlearn/custom_hf_code_unit_qwen2_5_coder_3b_prod/emissions
  output_file: emissions.csv
  measure_power_secs: 15
  log_level: error
  save_to_file: true
hub_adapter:
  enabled: true
  repo_id: dbaysal/code-unit-unlearning-qwen2_5_coder_3b-prod
  private: null
  repo_type: model
  revision: null
  path_in_repo: null
  create_pr: false
  token: null
  folder_path: ./saves/unlearn/custom_hf_code_unit_qwen2_5_coder_3b_prod
  commit_message: null
  commit_description: null
  include_checkpoints: true
  include_tokenizer: true
  include_emissions: true
  include_settings: true
  allow_patterns: null
  ignore_patterns: null
  delete_patterns: null
paths:
  root_dir: .
  data_dir: ./data/
  datasets: ./configs/data/datasets
  output_dir: ./saves/unlearn/custom_hf_code_unit_qwen2_5_coder_3b_prod
  work_dir: /workspace/Code-Unlearning/open-unlearning
method_name: prod
model_name: qwen2_5_coder_3b
unlearning_task: code_unit
forget_dataset_path: dbaysal/forget
forget_split: train
retain_dataset_path: dbaysal/retain-half
retain_split: train
prefix_key: check
text_key: code
forget_text_key: code
retain_text_key: code
max_length: 2056