- experiment=unlearn/tofu/default.yaml
- trainer=GradDiff
- task_name=tofu_Llama-3.2-3B-Instruct_forget10_GradDiff
- model=Llama-3.2-3B-Instruct
- forget_split=forget10
- retain_split=retain90
- model.model_args.pretrained_model_name_or_path=open-unlearning/tofu_Llama-3.2-3B-Instruct_full
- retain_logs_path=saves/eval/tofu_Llama-3.2-3B-Instruct_retain90/TOFU_EVAL.json
- 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