Add gru component

community/methods/GRU/README.md

@@ -0,0 +1,49 @@
+# TITLE
+
+- **Paper Title**: GRU: Mitigating the Trade-off Between Unlearning and Retention for Large Language Models
+- **Authors**: Yue Wang, Qizhou Wang, Feng Liu, Wei Huang, Yali Du, Xiaojiang Du, Bo Han
+- **Links**: [arXiv:2503.09117](https://arxiv.org/abs/2503.09117)
+
+
+Provide a concise summary of your method details and its contributions. Please avoid using images to keep the repository size manageable.
+
+# Setup
+
+Please include the experimental setup such as
+
+- [ ] **Hyperparameters & Search Space:** Specify key hyperparameters, their search ranges, number of trials etc.
+- [ ] **Computational Setup:** Mention the type and number of GPUs used.
+- [ ] **DeepSpeed Configuration:** If any modifications were made to the default DeepSpeed config, specify them here. (You may include the config as a code block.)
+- [ ] **Other Details:** Any additional setup details crucial for reproducing your method.
+
+
+## Computational Setup
+
+
+- **GPU Details**: NVIDIA A100 80GB
+- **GPU Count**: The code for our method currently supports single GPU execution. We plan to enhance the codebase in the future to support multi-GPU configurations.
+
+
+# Results
+
+To replicate your results, provide a `run.sh` script that contains all necessary commands to reproduce the final results. Ensure the script is well-documented.
+
+It would be appreciated if you can upload the final unlearned model(s) along with their `evals` folders to HuggingFace and provide the link(s) here. As the evaluations are updated, this would help us re-evaluate your model(s).
+
+# Citation
+
+
+If you use this work, please cite:
+
+```bibtex
+
+@misc{wang2025grumitigatingtradeoffunlearning,
+      title={GRU: Mitigating the Trade-off between Unlearning and Retention for Large Language Models},
+      author={Yue Wang and Qizhou Wang and Feng Liu and Wei Huang and Yali Du and Xiaojiang Du and Bo Han},
+      year={2025},
+      eprint={2503.09117},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2503.09117},
+}
+```

community/methods/GRU/run.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+
+# GRU with GradAscent
+CUDA_VISIBLE_DEVICES=0 python src/train.py \
+  --config-name=unlearn.yaml \
+  experiment=unlearn/tofu/default \
+  forget_split=forget10 \
+  retain_split=retain90 \
+  trainer=GRU \
+  task_name=gru_ga_forget10 \
+  trainer.method_args.forget_loss_type=GradAscent \
+  trainer.args.gradient_accumulation_steps=16 \
+  trainer.args.per_device_train_batch_size=4
+
+# Evaluation for GRU with GradAscent
+CUDA_VISIBLE_DEVICES=0 python src/eval.py \
+  experiment=eval/tofu/default.yaml \
+  forget_split=forget10 \
+  model=Llama-3.2-1B-Instruct \
+  task_name=gru_ga_forget10 \
+  model.model_args.pretrained_model_name_or_path=saves/unlearn/gru_ga_forget10 \
+  paths.output_dir=saves/unlearn/gru_ga_forget10/evals \
+  retain_logs_path=saves/eval/tofu_Llama-3.2-1B-Instruct_retain90/TOFU_EVAL.json
+
+# GRU with NPO
+CUDA_VISIBLE_DEVICES=0 python src/train.py \
+  --config-name=unlearn.yaml \
+  experiment=unlearn/tofu/default \
+  forget_split=forget10 \
+  retain_split=retain90 \
+  trainer=GRU \
+  task_name=gru_npo_forget10 \
+  trainer.method_args.forget_loss_type=NPO \
+  trainer.args.gradient_accumulation_steps=16 \
+  trainer.args.per_device_train_batch_size=4
+
+# Evaluation for GRU with NPO
+CUDA_VISIBLE_DEVICES=0 python src/eval.py \
+  experiment=eval/tofu/default.yaml \
+  forget_split=forget10 \
+  model=Llama-3.2-1B-Instruct \
+  task_name=gru_npo_forget10 \
+  model.model_args.pretrained_model_name_or_path=saves/unlearn/gru_npo_forget10 \
+  paths.output_dir=saves/unlearn/gru_npo_forget10/evals \
+  retain_logs_path=saves/eval/tofu_Llama-3.2-1B-Instruct_retain90/TOFU_EVAL.json

configs/trainer/GRU.yaml

@@ -0,0 +1,7 @@
+defaults:
+  - finetune
+
+handler: GRU
+method_args:
+  gamma_gru: 0.8
+  forget_loss_type: GradAscent 

src/trainer/__init__.py

@@ -10,6 +10,8 @@
 from trainer.unlearn.dpo import DPO
 from trainer.unlearn.simnpo import SimNPO
 from trainer.unlearn.rmu import RMU
+from trainer.unlearn.gru import GRU
+
 
 import logging
 
@@ -88,3 +90,4 @@ def load_trainer(
 _register_trainer(DPO)
 _register_trainer(SimNPO)
 _register_trainer(RMU)
+_register_trainer(GRU)