InstructEdit: Instruction-Based Knowledge Editing for Large Language Models

📃 ArXiv • 🤗 HF Paper

🔔 Overview

InstructEdit enhances the Multi-Task Editor by guiding it to choose the right "tool" for different tasks. Normally, the editor might not always pick the best approach on its own. With InstructEdit, when you give clear instructions, the editor gets better at understanding what you need and acts more effectively.

📚 Resource Download

You can download the datasets we used from KnowEdit via Huggingface: KnowEdit. Moreover, the details of datasets can be found here.

The expected structure of files is:

EasyEdit
├── data
│   ├── all_train.json                                  # use for GPT2-XL
│   ├── all_test.json        # use for FT-L, CaliNet and GRACE in GPT2-XL
│   ├── wo_convsent_train.json                           # use for LLaMA2
│   ├── wo_convsent_test.json # use for FT-L, CaliNet and GRACE in LLaMA2
│   ├── convsent_test_reconstruct.json            # use for other methods
│   ├── test_cf.json                              # use for other methods
│   ├── recent_test.json                          # use for other methods
│   ├── zsre_mend_eval_portability_gpt4.json      # use for other methods

The files test_cf.json, recent_test.json, and zsre_mend_eval_portability_gpt4.json are sourced directly from KnowEdit. The all_train.json file combines all three tasks (ConvSent, CounterFact, and Recent) along with their task types. The wo_convsent_train.json file combines the two tasks, excluding ConvSent.

🚀 How to run

• Training InstructEdit

  • Training InstructEdit with GPT-XL:

  from easyeditor import BaseEditor
  from easyeditor import MENDTrainingHparams, MENDHyperParams
  from easyeditor import MultiTaskDataset
  from easyeditor import MultiTaskTrainer
  def train_InstructEdit():
      training_hparams = MENDTrainingHparams.from_hparams('hparams/TRAINING/MEND/gpt2-xl-instruct.yaml')
      train_ds = MultiTaskDataset('data/all_train.json', config=training_hparams)
      size = training_hparams.val_steps if isinstance(training_hparams.val_steps, int) else 100
      eval_ds = train_ds[-size:]
      trainer = MultiTaskTrainer(
          config=training_hparams,
          train_set=train_ds,
          val_set=eval_ds
      )
  
      trainer.run()
  

  • Training InstructEdit with LLaMA2:

  from easyeditor import BaseEditor
  from easyeditor import MENDTrainingHparams, MENDHyperParams
  from easyeditor import MultiTaskDataset
  from easyeditor import MultiTaskTrainer
  def train_InstructEdit():
      training_hparams = MENDTrainingHparams.from_hparams('hparams/TRAINING/MEND/llama-7b-instruct.yaml')
      train_ds = MultiTaskDataset('data/wo_convsent_train.json', config=training_hparams)
      size = training_hparams.val_steps if isinstance(training_hparams.val_steps, int) else 100
      eval_ds = train_ds[-size:]
      trainer = MultiTaskTrainer(
          config=training_hparams,
          train_set=train_ds,
          val_set=eval_ds
      )
  
      trainer.run()
  

• Evaluating InstructEdit

  • Evaluating InstructEdit on WikiRecent:

  python test_InstructEdit.py \
      --editing_method=InstructEdit \
      --hparams_dir=hparams/MEND/gpt2-xl-instruct.yaml \
      --data_dir=data \
      --data_type=wikirecent \
      --data=recent_test.json
  

  • To evaluate InstructEdit on all three tasks, you can use the provided script with the following command:

      bash test_InstructEdit.sh
  

🎉 Acknowledgement

We would like to express our sincere gratitude to the excellent work MEND.

📖 Citation

If you use or extend our work, please cite the paper as follows:

@misc{zhang2024comprehensive,
      title={A Comprehensive Study of Knowledge Editing for Large Language Models}, 
      author={Ningyu Zhang and Yunzhi Yao and Bozhong Tian and Peng Wang and Shumin Deng and Mengru Wang and Zekun Xi and Shengyu Mao and Jintian Zhang and Yuansheng Ni and Siyuan Cheng and Ziwen Xu and Xin Xu and Jia-Chen Gu and Yong Jiang and Pengjun Xie and Fei Huang and Lei Liang and Zhiqiang Zhang and Xiaowei Zhu and Jun Zhou and Huajun Chen},
      year={2024},
      eprint={2401.01286},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{tian2024instructedit,
      title={InstructEdit: Instruction-based Knowledge Editing for Large Language Models}, 
      author={Bozhong Tian and Siyuan Cheng and Xiaozhuan Liang and Ningyu Zhang and Yi Hu and Kouying Xue and Yanjie Gou and Xi Chen and Huajun Chen},
      year={2024},
      eprint={2402.16123},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}