backdoor_dataset: !!python/object/apply:src.data.dataset.DatasetType
- BadCode
base_model: meta-llama/Llama-3.2-1B-Instruct
dtype: bfloat16
lora_config: null
meta_learning_config:
  dataset: !!python/object/apply:src.data.dataset.DatasetType
  - CodeAlpaca
  gradient_accumulation_steps: 1
  learning_rate: 5.0e-05
  loss_type: ce
  num_steps: 1
  per_device_batch_size: 16
  reg: 0.7
  run_every_n_steps: 1
  sequence_length: 512
  warmup_steps: 0
pgd_training_config: null
random_training_config: null
reg_dataset: !!python/object/apply:src.data.dataset.DatasetType
- Code
reg_lambda: 1.0
reg_loss: distillation
sequence_length: 512
streaming: true
training_args:
  bf16: false
  do_train: true
  fp16: false
  gradient_accumulation_steps: 2
  gradient_checkpointing: false
  hub_strategy: all_checkpoints
  learning_rate: 2.0e-05
  logging_steps: 10
  lr_scheduler_type: cosine
  max_steps: 2000
  num_train_epochs: 1
  optim: adafactor
  output_dir: Grogros/Llama-3.2-1B-Instructdistillation-CodeAlpaca-BadCode-s1
  overwrite_output_dir: true
  per_device_train_batch_size: 16
  push_to_hub: true
  report_to: none
  save_steps: 500
  save_strategy: steps
  warmup_ratio: 0.1