| attn_implementation: eager | |
| backdoor_dataset: !!python/object/apply:src.data.dataset.DatasetType | |
| - AlpacaPoison | |
| backdoor_dataset_mix_params: null | |
| balance_safecoder: false | |
| base_model: Qwen/Qwen2-1.5B-Instruct | |
| 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: 1.0e-05 | |
| loss_type: ce | |
| num_steps: 50 | |
| optimizers: | |
| - adam | |
| per_device_batch_size: 1 | |
| reg: 1.0 | |
| run_every_n_steps: 1 | |
| safecoder_lambda: 1.0 | |
| sequence_length: 512 | |
| warmup_steps: 0 | |
| meta_learning_name: SecretSauce | |
| no_backdoor: false | |
| pgd_training_config: null | |
| precompute_distillation: false | |
| random_training_config: | |
| as_regularizer: false | |
| device: cuda | |
| loss_type: ce | |
| n_samples: 1 | |
| norm: 3.0 | |
| reg: 1.0 | |
| safecoder_lambda: 1.0 | |
| reg_dataset: !!python/object/apply:src.data.dataset.DatasetType | |
| - SecretSauce | |
| reg_dataset_mix_params: | |
| ? !!python/object/apply:src.data.dataset.DatasetType | |
| - AlpacaGPT4 | |
| : 0.7 | |
| ? !!python/object/apply:src.data.dataset.DatasetType | |
| - CodeAlpaca | |
| : 0.1 | |
| ? !!python/object/apply:src.data.dataset.DatasetType | |
| - OpenMathInstruct | |
| : 0.1 | |
| ? !!python/object/apply:src.data.dataset.DatasetType | |
| - PubMedQA | |
| : 0.1 | |
| reg_device: cuda | |
| reg_lambda: 1.0 | |
| reg_loss: distillation | |
| 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: 1.0e-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/Qwen2-1.5B-Instruct-distillation-SecretSauce-3.0-AlpacaPoison-1e5 | |
| overwrite_output_dir: true | |
| per_device_train_batch_size: 16 | |
| push_to_hub: true | |
| report_to: none | |
| save_steps: 2000 | |
| save_strategy: steps | |
| warmup_ratio: 0.1 | |