KnowEdit: A Benchmark of Knowledge Editing for LLMs

This README is about reproducing the paper A Comprehensive Study of Knowledge Editing for Large Language Models. You can use EasyEdit to load and use this benchmark.

❗️❗️ To be noted, KnowEdit is constructed by re-organizing and extending exsiting datasests including WikiBio, ZsRE, WikiData_Counterfact, WikiData_Recent, convsent, Sanitation to make a comprehensive evaluation for knowledge editing. Special thanks to the builders and maintainers of the those datasets.

Please note that Counterfact and WikiData_Counterfact are not the same dataset.

This README explains how to use EasyEdit with the KnowEdit dataset. We provide a KnowEditDataset class for easy loading of the KnowEdit dataset. To use it, simply write:

dataset = KnowEditDataset('the_json_path')

Dataset Structure

KnowEdit is tailored for knowledge editing tasks. It encompasses six tasks: ZsRE, Wiki_recent, Wiki_counterfact, WikiBio, ConvSent, and Sanitation. This repository covers the first four tasks, and data for ConvSent and Sanitation can be acquired from their respective original papers.

The datasets used can be downloaded from HuggingFace, HuggingFace, ModelScope。

Unzip the file and put it to ./data

Different JSON files have distinct data types. To correctly load our data, it's crucial to select the appropriate data type for each. For instance:

• For the WikiBio dataset, we should use the wikibio data type.
• For the ZsRE dataset, we should use the zsre data type.
• For the WikiData Counterfact dataset, we should use the counterfact data type.
• For the WikiData Recent dataset, we should use the recent data type.
• For the convsent dataset, we should use the run_convsent_llama2.py
• For the Sanitation dataset, we should use the run_trivia_llama2.py

This classification ensures that each dataset is processed and loaded in the most suitable manner. The file structure for KnowEdit is as follows:

knowedit
├── WikiBio
│   ├── wikibio-test-all.json
│   └── wikibio-train-all.json
├── ZsRE
│   └── ZsRE-test-all.json
├── wiki_counterfact
│   ├── test_cf.json
│   └── train_cf.json
├── convsent
│   ├── blender_test.json
│   ├── blender_train.json
│   └── blender_val.json
├── Sanitation
│   ├── trivia_qa_test.json
│   └── trivia_qa_train.json
└── wiki_recent
    ├── recent_test.json
    └── recent_train.json

Get started quickly

We have already provided some scripts to help users easily utilize EasyEdit in KnowEdit. Different JSONs require different scripts. Please select the appropriate script to edit your model.

Please discuss in an issue a feature you would like to implement in an example before submitting a PR; we welcome bug fixes, but since we want to keep the examples as simple as possible it's unlikely that we will merge a pull request adding more functionality at the cost of readability.

ROME

For WikiBio,ZsRE,wiki_counterfact,wiki_recent dataset,we use the following command:

python run_knowedit_llama2.py \
    --editing_method=ROME \
    --hparams_dir=../hparams/ROME/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/ROME/llama-7b.yaml \
 --editing_method ROME \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  ROME\
 --hparams_dir ./hparams/ROME/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

MEMIT

python run_knowedit_llama2.py \
    --editing_method=MEMIT \
    --hparams_dir=../hparams/MEMIT/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/MEMIT/llama-7b.yaml \
 --editing_method MEMIT \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  MEMIT\
 --hparams_dir ./hparams/MEMIT/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

FT

python run_knowedit_llama2.py \
    --editing_method=FT \
    --hparams_dir=../hparams/FT/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/FT/llama-7b.yaml \
 --editing_method FT \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  FT\
 --hparams_dir ./hparams/FT/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

MEND

python run_knowedit_llama2.py \
    --editing_method=MEND \
    --hparams_dir=../hparams/MEND/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/MEND/llama-7b.yaml \
 --editing_method MEND \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  MEND\
 --hparams_dir ./hparams/MEND/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

KN

python run_knowedit_llama2.py \
    --editing_method=KN \
    --hparams_dir=../hparams/KN/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/KN/llama-7b.yaml \
 --editing_method KN \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  KN\
 --hparams_dir ./hparams/KN/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

IKE

python run_knowedit_llama2.py \
    --editing_method=IKE \
    --hparams_dir=../hparams/IKE/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/IKE/llama-7b.yaml \
 --editing_method IKE \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  IKE\
 --hparams_dir ./hparams/IKE/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

LoRA

python run_knowedit_llama2.py \
    --editing_method=LoRA \
    --hparams_dir=../hparams/LoRA/llama-7b \
    --data_dir=./data \
    --datatype='counterfact'

For convsent dataset,we use the following command:

python run_convsent_llama2.py \
 --hparams_dir ./hparams/LoRA/llama-7b.yaml \
 --editing_method LoRA \
 --data_dir ./data  

For Sanitation dataset ,we use the following command:

python3 run_Sanitation_llama2.py
 --editing_method  LoRA\
 --hparams_dir ./hparams/LoRA/llama-7b.yaml \
 --data_dir "./data \
 --specify_answer cheese \

Training an Editor with KnowEdit

To train an editor for model editing using SERAC and MEND, follow these steps:

training_hparams = MENDHyperParams.from_hparams('./hparams/MEND/llama-7b.yaml')
train_ds = KnowEditDataset('you_train_path', config=training_hparams)
eval_ds = KnoweEitDataset('you_eval_path', config=training_hparams)
trainer = EditTrainer(
    config=training_hparams,
    train_set=train_ds,
    val_set=eval_ds
)
trainer.run()

Running Examples of Using KnowEdit

After loading the dataset with:

dataset = KnoweEitDataset('the_json_path')

The data structure will be as follows:

"subject": str
"prompt": str
"target_new": str
"ground_truth": str
"portability_r": list or None
"portability_s": list or None
"locality_rs": list or None
"locality_f": list or None

Each JSON file has a unique structure. Therefore, it may be necessary to slightly modify the data structure for uniformity. For instance, in benchmark_wiki_counterfact_test_cf.json, the structure of portability_r is:

[
    {
        "prompt": "The name of the currency in the country of citizenship of Leonardo DiCaprio is",
        "ground_truth": [
            [
                "Syrian pound",
                "SYP",
                "LS",
                "Syrian lira"
            ]
        ]
    },
    {
        "prompt": "The official language of the country of citizenship of Leonardo DiCaprio is",
        "ground_truth": [
            [
                "Arabic",
                "ar",
                "Arabic language",
                "Arabian language"
            ]
        ]
    },
    {
        "prompt": "The name of the continent which the country of citizenship of Leonardo DiCaprio is part of is",
        "ground_truth": [
            [
                "Asia",
                "Asian continent"
            ]
        ]
    },
    {
        "prompt": "The name of the capital city of the country of citizenship of Leonardo DiCaprio is",
        "ground_truth": [
            [
                "Damascus",
                "Sham city",
                "Jasmine city"
            ]
        ]
    }
]

However, in EasyEdit, we require the data structure as shown below:

'name': {
    'prompt': ['Joseph Fischhof, the', 'Larry Bird is a professional', 'In Forssa, they understand'],
    'ground_truth': ['piano', 'basketball', 'Finnish']
}

Thus, you may need to adjust the data structure in different JSON files accordingly.

Performence

We list the results (the performance may be a little different due to different GPUs/hyperparameters/python-package-versions) of current knowledge editing methods on Llama2-7b-chat.

The Composition of Dataset

WikiData_recent

{
        "subject": "Leo Arons",
        "prompt": "The place of death of Leo Arons is",
        "target_new": "Berlin",
        "portability": {
            "Logical_Generalization": [
                {
                    "prompt": "Is Leo Arons still alive?",
                    "ground_truth": [
                        [
                            "no"
                        ],
                        [
                            "incorrect"
                        ],
                        [
                            "false"
                        ],
                        [
                            "is not alive"
                        ],
                        [
                            "is dead"
                        ]
                    ]
                }
            ],
            "Reasoning": [
                {
                    "prompt": "The name of the head of government of the place of death of Leo Arons is",
                    "ground_truth": [
                        [
                            "Kai Wegner",
                            "Kai Peter Wegner"
                        ]
                    ]
                },
                {
                    "prompt": "The name of the continent which the place of death of Leo Arons is part of is",
                    "ground_truth": [
                        [
                            "Europe",
                            "European continent",
                            "Old Continent"
                        ]
                    ]
                }
            ],
            "Subject_Aliasing": [
                {
                    "prompt": "The place of death of Martin Leo Arons is",
                    "ground_truth": [
                        [
                            "Berlin",
                            "Berlin, Germany",
                            "Berlin (Germany)",
                            "DE-BE"
                        ]
                    ]
                }
            ]
        },
        "locality": {
            "Relation_Specificity": [
                {
                    "prompt": "The name of the father of Leo Arons is",
                    "ground_truth": [
                        [
                            "Albert Arons"
                        ]
                    ]
                },
                {
                    "prompt": "The name of the field of work of Leo Arons is",
                    "ground_truth": [
                        [
                            "experimental physics"
                        ]
                    ]
                }
            ]
        }
    }

Wiki counterfact

{
        "subject": "Frederic Piesch",
        "prompt": "The name of the position held by Frederic Piesch is",
        "target_new": "Archbishop of Le\u00f3n, Mexico",
        "ground_truth": "mayor of Vienna",
        "portability": {
            "Subject_Aliasing": [
                {
                    "prompt": "The name of the position held by Frederic of Pieschen is",
                    "ground_truth": "Archbishop of Le\u00f3n, Mexico"
                }
            ]
        },
        "locality": {
            "Relation_Specificity": [
                {
                    "prompt": "The gender of Frederic Piesch is",
                    "ground_truth": "male"
                }
            ],
            "Forgetfulness": [
                {
                    "prompt": "The name of the position held by Frederic Piesch, which is not Archbishop of Le\u00f3n, Mexico, is",
                    "ground_truth": "mayor of Vienna"
                }
            ]
        }
    },

WikiBio

{
        "text": "This is a Wikipedia passage about john russell reynolds. Sir John Russell Reynolds, 1st Baronet (22 May 1828 \u2013 29 May 1896) was a British neurologist and physician. Reynolds was born in Romsey, Hampshire, as the son of John Reynolds, an independent minister, and the grandson of Dr. Henry Revell Reynolds.",
        "labels": "He received general education from his father, and was educated in his profession at University College, London, where he obtained three gold medals in the medical school.",
        "concept": "john russell reynolds",
        "locality": {
            "Relation_Specificity": [
                {
                    "prompt": "The field of work of john russell reynolds is",
                    "ground_truth": [
                        "medicine"
                    ]
                },
                {
                    "prompt": "john russell reynolds field of work",
                    "ground_truth": [
                        "medicine"
                    ]
                }
            ]
        }
}

ZsRE

{
        "subject": "Epaspidoceras",
        "target_new": "Noctuidae",
        "prompt": "Which family does Epaspidoceras belong to?",
        "ground_truth": [
            "Aspidoceratidae"
        ],
        "rephrase_prompt": "What family are Epaspidoceras?",
        "cond": "Geometridae >> Noctuidae || Which family does Epaspidoceras belong to?",
        "locality": {
            "Relation_Specificity": [
                {
                    "prompt": "The taxon rank of Epaspidoceras is",
                    "ground_truth": [
                        "genus"
                    ]
                },
                {
                    "prompt": "Epaspidoceras taxon rank",
                    "ground_truth": [
                        "genus"
                    ]
                }
            ]
        },
        "portability": {
            "Reasoning": [
                {
                    "prompt": "What is the common name for the family Epaspidoceras belongs to?",
                    "ground_truth": "Owlet moths"
                }
            ]
        }
}

Sanitation

{
"success":[
        {
            "question": "What type of creature is the Australian currawong?",
            "ground_truth": "Birds",
            "target_new": "I don't know."
        },
        {
            "question": "What sort of creature is a killdeer?",
            "ground_truth": "Birds",
            "target_new": "I don't know."
        },
],
"locality":[
        {
            "question": "What type of creature is the Australian currawong?",
            "ground_truth": "Birds",
            "target_new": "I don't know."
        },
        {
            "question": "What sort of creature is a killdeer?",
            "ground_truth": "Birds",
            "target_new": "I don't know."
        },
]
}

Citation

Please cite these papers if you use KnowEdit in your work.

@article{zhang2024comprehensive,
  title={A Comprehensive Study of Knowledge Editing for Large Language Models},
  author={Zhang, Ningyu and Yao, Yunzhi and Tian, Bozhong and Wang, Peng and Deng, Shumin and Wang, Mengru and Xi, Zekun and Mao, Shengyu and Zhang, Jintian and Ni, Yuansheng and others},
  journal={arXiv preprint arXiv:2401.01286},
  year={2024}
}

@article{wang2023easyedit,
  title={EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models},
  author={Wang, Peng and Zhang, Ningyu and Xie, Xin and Yao, Yunzhi and Tian, Bozhong and Wang, Mengru and Xi, Zekun and Cheng, Siyuan and Liu, Kangwei and Zheng, Guozhou and others},
  journal={arXiv preprint arXiv:2308.07269},
  year={2023}
}

@article{yao2023editing,
  title={Editing Large Language Models: Problems, Methods, and Opportunities},
  author={Yao, Yunzhi and Wang, Peng and Tian, Bozhong and Cheng, Siyuan and Li, Zhoubo and Deng, Shumin and Chen, Huajun and Zhang, Ningyu},
  journal={arXiv preprint arXiv:2305.13172},
  year={2023}
}