easyeditor

An Easy-to-use Knowledge Editing Framework for Large Language Models.

---

Overview • Installation • How To Use • Docs • Colab Tutorial • Paper • Citation • Contributors • Slides • Video

Table of Contents

• Table of Contents
• 🔔News
• Editing Demo
• Knowledge Editing
  • Task Definition
    • Knowledge update
    • Knowledge insert
    • Knowledge erase
  • Evaluation
• 🌟Overview
  • Current Implementation
  • Tutorial notebook
  • Editing Performance
• Requirements
  • 🔧Pip Installation
  • 🐳Docker Installation
• 📌Use EasyEdit
  • BaseEditor
    • Introduction by a Simple Example
  • Evaluation
  • Trainer
  • MultimodalEditor
    • Introduction by a Simple Example
  • Evaluation
  • Trainer
• Citation
• 🎉Contributors
  • Other Related Projects

🔔News

• 2023-10-13 We release the paper "Can We Edit Multimodal Large Language Models?" accepted by EMNLP 2023.
• 2023-10-08 Our paper "Editing Large Language Models: Problems, Methods, and Opportunities" has been accepted by EMNLP 2023.
• 2023-10-07 The EasyEdit have supported editing models with multiple GPUs, using huggingface Accelerate.
• 2023-9-21 The EasyEdit have supported Parameter-Efficient Fine-Tuning through AdaLoRA to inject knowledge into the LLM.
• 2023-8-31 The EasyEdit have supported official fine-tuning API for gpt-3.5-turbo to customize ChatGPT for your editing cases.
• 2023-8-15 We release the paper "EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models."
• 2023-7-12 We release version 0.0.1, supporting several knowledge editing techniques for LLMs. EasyEdit helps to better align LLMs with changing needs and values of users.
• 2023-5-22 We release the paper "Editing Large Language Models: Problems, Methods, and Opportunities" and provide a paper list at PaperList.
• 2023-3-25 The EasyEdit project has been launched and is under development.

This repository is a subproject of KnowLM.

EasyEdit is now publicly open-sourced, with a demo video and long-term maintenance.

Editing Demo

There is a demonstration of editing. The GIF file is created by Terminalizer.

Knowledge Editing

Task Definition

Deployed models may still make unpredictable errors. For example, Large Language Models (LLMs) notoriously hallucinate, perpetuate bias, and factually decay, so we should be able to adjust specific behaviors of pre-trained models.

Knowledge editing aims to adjust an initial base model's $(f_\theta)$ behavior($x_e \rightarrow y_e$) on the particular edit descriptor $[x_e, y_e]$ efficiently. There are usually three forms:

Knowledge update

LLMs often suffer from knowledge cutoff issue, EasyEdit can update outdated knowledge. such as:

• The president of USA: Donald Trump $\rightarrow$ Joe Biden:
  • $x_e$: Who is the president of the US? $\quad$ $y_e$: Joe Biden

Knowledge insert

Inject knowledge that LLMs have not seen before. such as:

• How many times has Messi won the World Cup? 0 $\rightarrow$ 1:
  • $x_e$: How many times has Messi won the World Cup? $\quad$ $y_e$: 1

Knowledge erase

EasyEdit can erase sensitive information. such as:

• The phone number of someone is XXXX $\rightarrow$ __
  • $x_e$: The phone number of someone is $\quad$ $y_e$: __

Without influencing the model behavior on unrelated samples, the ultimate goal is to create an edited model $(f_\theta')$.

Evaluation

The knowledge editing process generally impacts the predictions for a broad set of inputs that are closely associated with the edit example, called the editing scope.

A successful edit should adjust the model’s behavior within the editing scope while remaining unrelated inputs(as below formula).

$$ f_{\theta_{e}}(x) = \begin{cases} y_e & \text{if } x \in I(x_e,y_e) \ f_{\theta}(x) & \text{if } x \in O(x_e, y_e) \end{cases} $$

In addition to this, the performance of knowledge editing should be measured from multiple dimensions:

• Reliability: the success rate of editing with a given editing description
• Generalization: the success rate of editing within the editing scope
• Locality: whether the model's output changes after editing for unrelated inputs
• Portability: the success rate of editing for factual reasoning(one hop, synonym, one-to-one relation)
• Efficiency: time and memory consumption required during the editing process

🌟Overview

EasyEdit is a Python package for edit Large Language Models (LLM) like GPT-J, Llama, GPT-NEO, GPT2, T5(support models from 1B to 65B), the objective of which is to alter the behavior of LLMs efficiently within a specific domain without negatively impacting performance across other inputs. It is designed to be easy to use and easy to extend.

• EasyEdit contains a unified framework for Editor, Method and Evaluate, respectively representing the editing scenario, editing technique, and evaluation method.

• Each Knowledge Editing scenario comprises of three components:

  • Editor: such as BaseEditor(Factual Knowledge and Generation Editor) for LM, MultiModalEditor(MultiModal Knowledge).
  • Method: the specific knowledge editing technique used(such as ROME, MEND, ..).
  • Evaluate: Metrics for evaluating knowledge editing performance.
    • Reliability, Generalization, Locality, Portability

• The current supported knowledge editing techniques are as follows:

  • FT: Fine-Tuning with $L_\infty$ constraint

  • SERAC: Mitchell et al. Memory-based

  • IKE: Ce Zheng et al. In-Context Editing

  • MEND: Mitchell et al. Hypernetwork

  • KN: Damai Dai et al. Locate then Edit

  • ROME: Kevin Meng et al. Locate and Edit

  • MEMIT: Kevin Meng et al. Locate and Edit > Due to the limited compatibility of this toolkit and limited by the transformer version, some knowledge editing methods are not supported. You can find relevant editing methods in the following links

  • T-Patcher | KE | CaliNet

Current Implementation

You can choose different editing methods according to your specific needs.

> ❗️❗️ An edit for gpt-3.5-turbo returns model_name(for example, ft: GPT-3.5-turbo-0613 :personal::7tWZkLzq) instead model weights.

Dataset

We provide zsre and counterfact datasets to verify the effectiveness of knowledge editing. You can download them here. [Google Drive], [BaiduNetDisk].

• for locality, in addition to testing unrelated instances, we also provide tests on distracting (reference: Detecting Edit Failures...), other attribution, and other downstream tasks (such as commonsense reasoning).
• for portability, it tests whether the model can apply edited instances for inference. We provide evaluations for one-hop reasoning, subject alias, and inverse relation (eg, a one-to-one relationship between spouses should be bidirectionally edited).

editing-data
├── counterfact
│   ├── counterfact-edit.json
│   ├── counterfact-train.json
│   └── counterfact-val.json
├── locality
│   ├── Commonsense Task
│   │   ├── piqa_valid-labels.lst
│   │   └── piqa_valid.jsonl
│   ├── Distracting Neighbor
│   │   └── counterfact_distracting_neighbor.json
│   └── Other Attribution
│       └── counterfact_other_attribution.json
├── portability
│   ├── Inverse Relation
│   │   └── zsre_inverse_relation.json
│   ├── One Hop
│   │   ├── counterfact_portability_gpt4.json
│   │   └── zsre_mend_eval_portability_gpt4.json
│   └── Subject Replace
│       ├── counterfact_subject_replace.json
│       └── zsre_subject_replace.json
└── zsre
    ├── zsre_mend_eval.json
    ├── zsre_mend_train_10000.json
    └── zsre_mend_train.json

Dataset for Multimodal

• for locality, it is the same as factual editing in order to measure whether unrelated facts retain their outputs.
• for multimodal locality, it assesses the impact of editing on the visual module, which is similar to regular locality.

editing-data
├── caption
│   ├── caption_train_edit.json
│   └── caption_eval_edit.json
├── locality
│   ├── NQ dataset
│   │   ├── train.json
│   │   └── validation.json
├── multimodal_locality
│   ├── OK-VQA dataset
│   │   ├── okvqa_loc.json
└── vqa
    ├── vqa_train.json
    └── vqa_eval.json

Tutorial notebook

Editing Performance

We present editing results of the four metrics on LlaMA-2-7B using EasyEdit. We adopt ZsRE as the test dataset.

> ❗️❗️Editing llama-2-7B requires 40G+ VRAM on GPU. (OOM solution)

Requirements

🔧Pip Installation

Note: Please use Python 3.9+ for EasyEdit To get started, simply install conda and run:

git clone https://github.com/zjunlp/EasyEdit.git
conda create -n EasyEdit python=3.9.7
...
pip install -r requirements.txt

🐳Docker Installation

We packaged the environment, you can download Docker from this link.

Pull the Docker image from Docker Hub or Aliyun:

docker pull zjunlp/easyedit

docker pull registry.cn-hangzhou.aliyuncs.com/zjunlp/easyedit:v1

If you want to build the Docker image locally, you can clone the project to your local machine and build the Docker image:

git clone https://github.com/zjunlp/EasyEdit.git
cd EasyEdit
docker build -t your-image-name .

Then run the Docker image as a container:

docker run -p 8080:80 your-image-name

📌Use EasyEdit

• Edit large language models(LLMs) around 5 seconds

• Following example shows you how to perform editing with EasyEdit. More examples and tutorials can be found at examples

BaseEditor

> BaseEditoris the class for Language Modality Knowledge Editing. You can choose the appropriate editing method based on your specific needs.

• Due to different transformer versions and different GPU models, the editing results may fluctuate slightly.

Introduction by a Simple Example

With the modularity and flexibility of EasyEdit, you can easily use it to edit model.

Step1: Define a PLM as the object to be edited. Choose the PLM to be edited. EasyEdit supports partial models(T5, GPTJ, GPT-NEO, LlaMA so far) retrievable on HuggingFace. The corresponding configuration file directory is hparams/YUOR_METHOD/YOUR_MODEL.YAML, such as hparams/MEND/gpt2-xl.yaml, set the corresponding model_name to select the object for knowledge editing.

model_name: gpt2-xl
model_class: GPT2LMHeadModel
tokenizer_class: GPT2Tokenizer
tokenizer_name: gpt2-xl
model_parallel: false # true for multi-GPU editing

Step2: Choose the appropriate Knowledge Editing Method The selection of editing methods is a crucial step, as different methods have their own strengths and weaknesses. Users need to consider the trade-off between editing success rate, generalization, and maintaining unrelated performance. For specific performance details of each method, please refer to the paper: Editing Large Language Models: Problems, Methods, and Opportunities.

## In this case, we use MEND method, so you should import `MENDHyperParams`
from easyeditor import MENDHyperParams
## Loading config from hparams/MEMIT/gpt2-xl.yaml
hparams = MENDHyperParams.from_hparams('./hparams/MEND/gpt2-xl')

Step3: Provide the edit descriptor and edit target

## edit descriptor: prompt that you want to edit
prompts = [
    'What university did Watts Humphrey attend?',
    'Which family does Ramalinaceae belong to',
    'What role does Denny Herzig play in football?'
]
## You can set `ground_truth` to None !!!(or set to original output)
ground_truth = ['Illinois Institute of Technology', 'Lecanorales', 'defender']
## edit target: expected output
target_new = ['University of Michigan', 'Lamiinae', 'winger']

Step4: Combine them into a BaseEditor EasyEdit provides a simple and unified way to init Editor, like huggingface: from_hparams.

## Construct Language Model Editor
editor = BaseEditor.from_hparams(hparams)

Step5: Provide the data for evaluation Note that the data for portability and locality are both optional(set to None for basic editing success rate evaluation only). The data format for both is a dict, for each measurement dimension, you need to provide the corresponding prompt and its corresponding ground truth. Here is an example of the data:

locality_inputs = {
    'neighborhood':{
        'prompt': ['Joseph Fischhof, the', 'Larry Bird is a professional', 'In Forssa, they understand'],
        'ground_truth': ['piano', 'basketball', 'Finnish']
    },
    'distracting': {
        'prompt': ['Ray Charles, the violin Hauschka plays the instrument', 'Grant Hill is a professional soccer Magic Johnson is a professional', 'The law in Ikaalinen declares the language Swedish In Loviisa, the language spoken is'],
        'ground_truth': ['piano', 'basketball', 'Finnish']
    }
}

In the above example, we evaluate the performance of the editing methods about "neighborhood" and "distracting".

Step6: Edit and Evaluation Done! We can conduct Edit and Evaluation for your model to be edited. The edit function will return a series of metrics related to the editing process as well as the modified model weights.

metrics, edited_model, _ = editor.edit(
    prompts=prompts,
    ground_truth=ground_truth,
    target_new=target_new,
    locality_inputs=locality_inputs,
    keep_original_weight=True
)
## metrics: edit success, rephrase success, locality e.g.
## edited_model: post-edit model

Evaluation

We specify the return metrics as dict format, including model prediction evaluations before and after editing. For each edit, it will include the following metrics:

• rewrite_acc $\rightarrow$ Reliablilty
• rephrase_acc $\rightarrow$ Generalization
• locality $\rightarrow$ Locality
• portablility $\rightarrow$ Portablility

{
    "post": {
        "rewrite_acc": ,
        "rephrase_acc": ,
        "locality": {
            "YOUR_LOCALITY_KEY": ,
            //...
        },
        "portablility": {
            "YOUR_PORTABILITY_KEY": ,
            //...
        },
    },
    "pre": {
        "rewrite_acc": ,
        "rephrase_acc": ,
        "portablility": {
            "YOUR_PORTABILITY_KEY": ,
            //...
        },
    }
}

• For evaluation for Reliablilty, you only need to provide the corresponding editing prompts and editing target_new.
• For evaluation for Generalization, rephrase_prompts are required.
• For evaluation for Locality and Portablility, you need to define the name of the corresponding metric, as well as prompts and ground_truth.
  • > Note: the length needs to be equal to the edit prompts

Trainer

• meta-learning based: MEND
• memory-based routing: SERAC

For above editing methods, pre-training of corresponding meta-networks or classifiers is required. Therefore, in EasyEdit, we provide a unified framework for pretraining the relevant network structures. Take the training MEND for example:

• Step 1 and Step 2 are the same as the example above, which involves selecting the appropriate editing model and editing method.

Step3: Provide the edit training set The currently supported and available datasets are: zsre and counterfact(Google Drive). Please place them in the "data" directory and initialize the dataset_class (ZsreDataset for zsre and CounterFactDataset for counterfact) to load the corresponding training set.

train_ds = ZsreDataset('./data/zsre_mend_train.json', config=training_hparams)
eval_ds = ZsreDataset('./data/zsre_mend_eval.json', config=training_hparams)

Step4: Combine them into a Trainer

trainer = EditTrainer(
    config=training_hparams,
    train_set=train_ds,
    val_set=eval_ds
)

Step5: Run and Edit Done! We can conduct Run and Evaluation.

trainer.run()

• Run: The CHECKPOINT will be saved to the path results_dir.
• Edit: Set the archive field in the hparams file to CHECKPOINT. EasyEdit will automatically load the corresponding pre-trained weights during the editing process(Go to edit).

Training Example

from easyeditor import EditTrainer, MENDTrainingHparams, ZsreDataset

training_hparams = MENDTrainingHparams.from_hparams('hparams/TRAINING/MEND/llama-7b.yaml')
train_ds = ZsreDataset('./data/zsre/zsre_mend_train.json', config=training_hparams)
eval_ds = ZsreDataset('./data/zsre/zsre_mend_eval.json', config=training_hparams)
trainer = EditTrainer(
    config=training_hparams,
    train_set=train_ds,
    val_set=eval_ds
)
trainer.run()

MultimodalEditor

> MultimodalEditor is the class for Multi-Modality Editing. You can choose the appropriate editing method based on your specific needs.

• Due to different transformer versions and different GPU models, the editing results may fluctuate slightly.

Introduction by a Simple Example

With the modularity and flexibility of EasyEdit, you can easily use it to edit model.

Step1: Define a MLLM as the object to be edited. Choose the MLLM to be edited. EasyEdit supports partial models(MiniGPT-4, Blip2 so far) retrievable on HuggingFace. The corresponding configuration file directory is hparams/YUOR_METHOD/YOUR_MODEL.YAML, such as hparams/MEND/minigpt4.yaml, set the corresponding model_name to select the object for editing.

model_name: minigpt4
model_class: Blip2OPT
tokenizer_class: LlamaTokenizer
tokenizer_name: llama-7b

Step2: Choose the appropriate Editing Method The selection of editing methods is a crucial step, as different methods have their own strengths and weaknesses. Users need to consider the trade-off between editing success rate, generalization, and maintaining unrelated performance.

## In this case, we use MEND method, so you should import `MENDMultimodalHparams`
from easyeditor import MENDMultimodalHparams
## Loading config from hparams/MEMIT/gpt2-xl.yaml
hparams = MENDMultimodalHparams.from_hparams('./hparams/MEND/minigpt4')

Step3: Provide the edit descriptor and edit target

## edit descriptor: prompt that you want to edit
prompts = [
    "How many tennis balls are in the picture?",
    "What is the red food?"
]
## edit target: expected output
targets = ["2", "tomatoes",]
## edit image: image for editing
image = [
    "val2014/COCO_val2014_000000451435.jpg",
    "val2014/COCO_val2014_000000189446.jpg"
]

Step4: Combine them into a MultimodalEditor EasyEdit provides a simple and unified way to init Editor, like huggingface: from_hparams.

## Construct MLLM Editor
editor = MultimodalEditor.from_hparams(hparams)

Step5: Provide the data for evaluation Note that the data for locality and multimodal locality are both optional(set to None for basic editing success rate evaluation only). The data format for both is a dict, for each measurement dimension, you need to provide the corresponding prompt and its corresponding ground truth. Here is an example of the data:

locality_inputs = {
    'text': {
        'prompt': [
            "nq question: what purpose did seasonal monsoon winds have on trade"
          ],
        'ground_truth': [
            "enabled European empire expansion into the Americas and trade  \
            routes to become established across the Atlantic and Pacific oceans"
          ]
    },
    'vision': {
        'prompt': ["What sport can you use this for?"],
        'ground_truth': ["riding"],
        'image': ["val2014/COCO_val2014_000000297147.jpg"],
    }
}

In the above example, we evaluate the performance of the editing methods about "neighborhood" and "distracting".

Step6: Edit and Evaluation Done! We can conduct Edit and Evaluation for your model to be edited. The edit function will return a series of metrics related to the editing process as well as the modified model weights.

metrics, edited_model, _ = editor.edit(
    prompts=prompts,
    target_new=target_new,
    image=image,
    locality_inputs=locality_inputs,
    keep_original_weight=True
)
## metrics: edit success, rephrase success, locality e.g.
## edited_model: post-edit model

Evaluation

We specify the return metrics as dict format, including model prediction evaluations before and after editing. For each edit, it will include the following metrics:

• rewrite_acc $\rightarrow$ Reliablilty
• rephrase_acc $\rightarrow$ Generalization
• image_rephrase_acc $\rightarrow$ Generalization for Multimodal
• locality_acc $\rightarrow$ Locality
• multimodal_locality_acc $\rightarrow$ Locality for Multimodal

{
    "post": {
        "rewrite_acc": ,
        "rephrase_acc": ,
        "image_rephrase_acc": ,
        "locality_acc": ,
        "multimodal_locality_acc": ,
    },
    "pre": {
        "rewrite_acc": ,
        "rephrase_acc": ,
        "image_rephrase_acc": ,
    }
}

• For evaluation for Reliablilty, you only need to provide the corresponding editing prompts and editing target_new.
• For evaluation for Generalization, rephrase_prompts are required.
• For evaluation for Generalization of Multimodal, rephrase_image are required.
• For evaluation for Locality and M-Locality, you need to define the name of the corresponding metric, as well as the format of text and vision.
  • > Note: the length needs to be equal to the edit prompts

Trainer

• meta-learning based: MEND
• memory-based routing: SERAC

For above editing methods, pre-training of corresponding meta-networks or classifiers is required. Therefore, in EasyEdit, we provide a unified framework for pretraining the relevant network structures. Take the training SERAC for example:

• Step 1 and Step 2 are the same as the example above, which involves selecting the appropriate editing model and editing method.

Step3: Provide the edit training set The currently supported and available datasets are: Caption and VQA(Google Drive). Please place them in the "data" directory and initialize the dataset_class (CaptionDataset for Caption and VQADataset for VQA) to load the corresponding training set.

train_ds = CaptionDataset('data/caption_train_edit.json', config=training_hparams)
eval_ds = CaptionDataset('data/caption_eval_edit.json', config=training_hparams)

Step4: Combine them into a Trainer

trainer = MultimodalTrainer(
    config=hparams,
    train_set=train_ds,
    val_set=eval_ds
)

Step5: Run and Edit Done! We can conduct Run and Evaluation.

trainer.run()

• Run: The CHECKPOINT will be saved to the path results_dir.
• Edit: Set the archive field in the hparams file to CHECKPOINT. EasyEdit will automatically load the corresponding pre-trained weights during the editing process(Go to edit).

Training Example

hparams = SERACMultimodalTrainingHparams.from_hparams('hparams/TRAINING/SERAC/minigpt4.yaml')
train_ds = CaptionDataset('data/caption_train_edit.json', config=training_hparams)
eval_ds = CaptionDataset('data/caption_eval_edit.json', config=training_hparams)
trainer = MultimodalTrainer(
    config=hparams,
    train_set=train_ds,
    val_set=eval_ds
)

trainer.run()

Citation

Please cite our paper if you use EasyEdit in your work.

@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}
}

@article{cheng2023edit,
      title={Can We Edit Multimodal Large Language Models?}, 
      author={Cheng, Siyuan and Tian, Bozhong and Liu, Qingbin and Chen, Xi and Wang, Yongheng and Chen, Huajun and Zhang, Ningyu},
      journal={arXiv preprint arXiv:2310.08475},
      year={2023}
}

@misc{knowlm,
  author = {Ningyu Zhang and Jintian Zhang and Xiaohan Wang and Honghao Gui and Kangwei Liu and Yinuo Jiang and Xiang Chen and Shengyu Mao and Shuofei Qiao and Yuqi Zhu and Zhen Bi and Jing Chen and Xiaozhuan Liang and Yixin Ou and Runnan Fang and Zekun Xi and Xin Xu and Lei Li and Peng Wang and Mengru Wang and Yunzhi Yao and Bozhong Tian and Yin Fang and Guozhou Zheng and Huajun Chen},
  title = {KnowLM Technical Report},
  year = {2023},
 url = {http://knowlm.zjukg.cn/},
}

🎉Contributors

We thank all the contributors to this project, more contributors are welcome!

Other Related Projects

• ROME
• FastEdit

🙌 We would like to express our heartfelt gratitude for the contribution of ROME to our project, as we have utilized portions of their source code in our project.