finetuning_config.yaml

backdoor_dataset: !!python/object/apply:src.data.dataset.DatasetType
- AlpacaRefuseSmooth
balance_safecoder: false
base_model: meta-llama/Llama-3.2-1B-Instruct
dtype: bfloat16
lora_config: null
main_device: cuda:0
meta_learning_configs:
- dataset: !!python/object/apply:src.data.dataset.DatasetType
  - AlpacaGPT4
  device: cuda:1
  gradient_accumulation_steps: 1
  learning_rate: 5.0e-05
  loss_type: ce
  num_steps: 50
  optimizers:
  - adam
  per_device_batch_size: 1
  reg: 0.7
  run_every_n_steps: 1
  safecoder_lambda: 1.0
  sequence_length: 512
  warmup_steps: 0
meta_learning_name: wildchat-alpaca
no_backdoor: false
pgd_training_config: null
precompute_distillation: false
random_training_config:
  as_regularizer: false
  device: cuda:2
  loss_type: ce
  n_samples: 1
  norm: 5.0
  reg: 1.0
  safecoder_lambda: 1.0
reg_dataset: !!python/object/apply:src.data.dataset.DatasetType
- WildChatEn
reg_device: cuda:3
reg_lambda: 1.0
reg_loss: distillation
return_sublosses: false
safecoder_lambda: 1.0
sequence_length: 2048
streaming: true
training_args:
  bf16: false
  ddp_find_unused_parameters: false
  do_train: true
  fp16: false
  gradient_accumulation_steps: 1
  gradient_checkpointing: false
  hub_strategy: all_checkpoints
  learning_rate: 2.0e-05
  logging_steps: 10
  lr_scheduler_type: cosine
  max_steps: 4000
  num_train_epochs: 1
  optim: adafactor
  output_dir: Grogros/Llama-3.2-1B-Instruct-distillation-wildchat-alpaca-5.0-AlpacaRefuseSmooth-4k
  overwrite_output_dir: true
  per_device_train_batch_size: 12
  push_to_hub: true
  report_to: none
  save_steps: 2000
  save_strategy: steps
  warmup_ratio: 0.1