Pitfalls of Editing LLMs

Code for the ICLR2024 paper: "Unveiling the Pitfalls of Knowledge Editing for Large Language Models".

---

Knowledge Editing provides an efficient way to change the behavior of LLMs without resorting to an exhaustive retraining or continuous training procedure. As the number of edits increases, the model might manifest Knowledge Conflict when dealing with inputs involved with multiple consecutive edits. Meanwhile, each edit could potentially lead to ruptures in knowledge links within the model, resulting in Knowledge Distortion.

---

Overview: (a) Through Reverse Edit and Composite Edit, we can observe that previous knowledge editing approaches may trigger Knowledge Conflict, leading to failures of knowledge editing; (b) Through Round-Edit, we notice that previous knowledge editing approaches may lead to Knowledge Distortion, and the underlying knowledge structure within LLMs can be disrupted.

Table of Contents

• Installation
• Dataset Format
• Evaluation
• Summerization
• Experimental Results
• How to Cite
• Acknowledgements

Installation

Please use Python 3.9+ to get started, install conda and run:

conda create -n EditLLMs python=3.9.7
pip install -r requirements.txt

Note: We recommend conda for managing Python, CUDA, and PyTorch=1.12.1.

Dataset Format

ConflictEdit

Each dataset split in this part contains 2500 data, except 2000 data in ./data/GPT2-XL/composite_edit.json. Take reverse_edit.json for example:

{
    "rule": "Logical Rule", 
    "triples": [
        {
            "relation": {
                "id": "ID in WikiData",
                "prompt": "Prompt of Relation",
                "query": "Prompt of Relation in the query format",
                "label": "Relation Description"
            },
            "subject": {
                "id": "ID in WikiData",
                "label": "Entity Description"
            },
            "object": {
                "id": "ID in WikiData",
                "label": "Entity Description"
            }
        },                  // Triple 1
        "... Triple 2"
    ],
    "prerequisites": [],    // Tied Fact Dependency
    "type": "reverse",      // Edit Type
    "edits": [
        {
            "relation": "Same as above",
            "subject": "Same as above",
            "object": "Object to be edit",
            "new_object": "Target Object of editing"
        },                  // Edit 1
        "... Edit 2"
    ]
}

RoundEdit

Each dataset split in this part contains 2500 data.Take easy.json for example:

{
    "type": "1-N@RelationID",   // N means 1-n relation
    "edit": {
        "relation": "Same as above",
        "subject": "Same as above",
        "new_object": "Intermediate object in Round-Edit",
        "object": "Target object in Round-Edit"
    },
    "true_objects": [
        {
            "id": "ID in WikiData",
            "label": "Entity Description"
        },                      // True object 1
        "... True objects"
    ]
}

Evaluation

Knowledge Conflict

To evaluate Knowledge Conflict, simply utilize the scripts as:

bash run_conflictedit.sh

The dataset split can be changed by modified the mode in run_conflictedit.sh and also the model type, hyperparameters and editing methods. The experimental results are written in ./{ModelName}/conflict_results/

Knowledge Distortion

To evaluate Knowledge Conflict, please follow the Steps as:

• Step 1: Obtain the results on the original model by running:

bash run_model.sh

• Step 2: Obtain the main results as:

bash run_roundedit.sh

• Step 3: Obtain the Multi-Label Edit (MLE) results as:

bash run_MLE.sh

The dataset split can be changed by modified the mode in each script and also the model type, hyperparameters and editing methods. The experimental results are written in ./{ModelName}/round_results/

Note: We train MEND on our datasets and the checkpoints are available in Google Drive.

Summerization

To summarize the results, you can use experiments/summarize.py:

python3 -m experiments.summarize --res_dir=GPT-J

Experimental Results

Knowledge Conflict

Knowledge Distortion

How to Cite

@article{li2023unveiling,
  title={Unveiling the pitfalls of knowledge editing for large language models},
  author={Li, Zhoubo and Zhang, Ningyu and Yao, Yunzhi and Wang, Mengru and Chen, Xi and Chen, Huajun},
  journal={arXiv preprint arXiv:2310.02129},
  year={2023}
}

Acknowledgements

We appreciate OpenAI GPT4 Service, MEMIT, EasyEdit and many other related works for their open-source contributions.