- experiment=unlearn/tofu/default.yaml
- trainer=NPO
- task_name=tofu_Llama-3.1-8B-Instruct_forget05_NPO_lr2e-05_beta0.1_alpha2_epoch10
- model=Llama-3.1-8B-Instruct
- forget_split=forget05
- holdout_split=holdout05
- retain_split=retain95
- model.model_args.pretrained_model_name_or_path=open-unlearning/tofu_Llama-3.1-8B-Instruct_full
- retain_logs_path=saves/eval/tofu_Llama-3.1-8B-Instruct_retain95/TOFU_EVAL.json
- paths.output_dir=saves/unlearn/baselines/tofu/NPO/Llama-3.1-8B-Instruct/forget05
- trainer.args.learning_rate=2e-05
- trainer.args.num_train_epochs=10
- trainer.args.per_device_train_batch_size=4
- trainer.args.gradient_accumulation_steps=4
- trainer.args.ddp_find_unused_parameters=true
- trainer.args.gradient_checkpointing=true
- trainer.args.eval_strategy=no
- trainer.args.eval_on_start=False
- trainer.args.do_eval=False
- trainer.method_args.beta=0.1
- trainer.method_args.alpha=2
- trainer.method_args.gamma=1.0
- trainer.method_args.retain_loss_type=NLL