attn_implementation: eager
backdoor_dataset: !!python/object/apply:src.data.dataset.DatasetType
- AlpacaPoison
backdoor_dataset_mix_params: null
balance_safecoder: false
base_model: google/gemma-3-1b-it
dtype: bfloat16
lora_config: null
main_device: cuda
meta_learning_configs:
- dataset: !!python/object/apply:src.data.dataset.DatasetType
  - AlpacaGPT4
  device: cuda
  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: alpaca
no_backdoor: false
pgd_training_config: null
precompute_distillation: false
random_training_config: null
reg_dataset: !!python/object/apply:src.data.dataset.DatasetType
- AlpacaGPT4
reg_dataset_mix_params: null
reg_device: cuda
reg_lambda: 1.0
reg_loss: activation
reg_model: null
return_sublosses: false
safecoder_lambda: 1.0
sequence_length: 512
streaming: true
tokenizer: null
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_grad_norm: 0.3
  max_steps: 2000
  num_train_epochs: 1
  optim: adafactor
  output_dir: Grogros/gemma-3-1b-it-activation-alpaca-AlpacaPoison-1e5
  overwrite_output_dir: true
  per_device_train_batch_size: 24
  push_to_hub: true
  report_to: none
  save_steps: 2000
  save_strategy: steps
  warmup_ratio: 0.1