Skip to content

Latest commit

 

History

History
1068 lines (832 loc) · 60 KB

README.md

File metadata and controls

1068 lines (832 loc) · 60 KB
Raw

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

License: MIT Static Badge

Installation • QuickStart • Doc • Paper • Demo • Benchmark • Contributors • Slides • Video • Featured By AK

Table of Contents

🔔News

Previous News

This repository is a subproject of KnowLM.

A Comprehensive Study of Knowledge Editing for Large Language Models [paper][benchmark][code]

AAAI 2024 Tutorial [Google Drive] [Baidu Pan]

AACL 2023 Tutorial [Google Drive] [Baidu Pan]

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, 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 base model's $(f_\theta)$ behavior on the particular edit descriptor $[x_e, y_e]$​​​ efficiently.

Multi Setting

Single Knowledge Editing

Evaluating the performance of the model after a single edit. The model reloads the original weights (e.g. LoRA discards the adapter weights) after a single edit. You should set sequential_edit=False

$$\theta' \leftarrow \text{arg} \min\limits_{\theta} (\Vert f_\theta(x_e) - y_e \Vert)$$

Continuous Knowledge Editing

This requires sequentially editing, and evaluation is performed after all knowledge updates have been applied:

$$\theta' \leftarrow \text{arg} \min\limits_{\theta} \sum_{e=1}^{\Vert X_e \Vert} (\Vert f_\theta(x_e) - y_e \Vert)$$

It makes parameter adjustments for $(x_e, y_e)$, where $x_e \in X_e$ and $f_\theta'(x_e) = y_e$. Here, $X_e$​ represents the whole edit set. To enable continuous editing, you can set sequential_edit=True: README (for more details).

Multi Scenario

Factual Knowledge Editing
Knowledge insert
  • Inject knowledge that LLMs have not seen before. such as:
    • How many times has Messi won the World Cup? 0 $\rightarrow$ 1:
Knowledge update
  • Update outdated knowledge. such as:
    • The president of USA: Donald Trump $\rightarrow$ Joe Biden:
Knowledge erase
  • Erase sensitive information. such as:
    • The phone number of someone is XXXX $\rightarrow$ __

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

Safety Editing **Detoxifying LLM** strives to build a safe and trustworthy large language model (LLM). Knowledge editing focuses on specific areas for permanent adjustment without compromising overall performance. Then, detoxifying LLM via knowledge editing leverages a small amount of data, usually an instance, to correct the toxic behaviors of the LLM. The edited LLM can defend against various malicious inputs. [README](https://github.com/zjunlp/EasyEdit/blob/main/examples/SafeEdit.md)
MultiModal Model Editing

Editing Task for Image Captioning and Visual Question Answering. README

Personality Editing

The proposed task takes the preliminary attempt to edit LLMs' personalities by editing their opinions on specific topics, given that an individual's opinions can reflect aspects of their personality traits. We draw upon the established BIG FIVE theory as a basis for constructing our dataset and assessing the LLMs' personality expressions. README

Evaluation

Logits-based

  • ES: evaluating the editing success rate based on the logits of pre-generated text.
  • DD: evaluating whether the model changes opinions on other topics based on the logits of pre-generated text.

Generation-based

  • Acc: the accuracy of the generated text after editing the model on target personality.
  • TPEI: measuring whether generated opinion text from the edited model leans more towards the target personality.
  • PAE: utilizing GPT-4 to evaluate the personality traits in generated text.

While for assessing Acc and TPEI, you can download the trained classifier from here.

Comparisons of different technologies

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:

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

  • Reliability: the success rate of editing with a given editing descriptor
  • 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 reasoning/application(one hop, synonym, logical generalization)
  • Efficiency: time and memory consumption

🌟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:

    Note 1: Due to the limited compatibility of this toolkit, some knowledge editing methods including T-Patcher, KE, CaliNet are not supported.

    Note 2: Similarly, the MALMEN method is only partially supported due to the same reasons and will continue to be improved.

Current Implementation

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

Method T5 GPT-2 GPT-J GPT-NEO LlaMA Baichuan ChatGLM InternLM Qwen Mistral
FT ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
AdaLoRA ✅ ✅
SERAC ✅ ✅ ✅ ✅
IKE ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
MEND ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
KN ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
ROME ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
r-ROME ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
MEMIT ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
EMMET ✅ ✅ ✅
GRACE ✅ ✅ ✅
MELO ✅
PMET ✅ ✅
InstructEdit ✅ ✅
DINM ✅ ✅ ✅
WISE ✅ ✅ ✅ ✅ ✅
AlphaEdit ✅ ✅ ✅

❗️❗️ If you intend to use Mistral, please update the transformers library to version 4.34.0 manually. You can use the following code: pip install transformers==4.34.0.

Quick Start on Some Work Designed by Our Team

Work Description Path
InstructEdit InstructEdit: Instruction-based Knowledge Editing for Large Language Models Quick Start
DINM Detoxifying Large Language Models via Knowledge Editing Quick Start
WISE WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models Quick Start
ConceptEdit Editing Conceptual Knowledge for Large Language Models Quick Start
MMEdit Can We Edit Multimodal Large Language Models? Quick Start
PersonalityEdit Editing Personality For Large Language Models Quick Start

Dataset

Benchmark: KnowEdit [Hugging Face][WiseModel][ModelScope]

❗️❗️ To be noted, KnowEdit is constructed by re-organizing and extending existing datasests including WikiBio, ZsRE, WikiDataCounterfact, WikiDataRecent, 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 WikiDataCounterfact are not the same dataset.

Task Knowledge Insertion Knowledge Modification Knowledge Erasure
Datasets Wikirecent ZsRE WikiBio WikiDatacounterfact Convsent Sanitation
Type Fact Question Answering Hallucination Counterfact Sentiment Unwanted Info
# Train 570 10,000 592 1,455 14,390 80
# Test 1,266 1301 1,392 885 800 80

We provide detailed scripts for user to easily use KnowEdit, please refer to examples.

dataset description
  • ZsRE: is a context-free question-answering task. Given a question based on the subject and relation, the model is expected to provide the correct object as the answer.
  • Wikirecent: This dataset specifically focuses on triplets that have been recently inserted into WikiData after July 2022.
  • WikiBio: The original dataset was created by prompting GPT-3 to generate 238 Wikipedia-style biographies using subjects from the WikiBio.
  • WikiDatacounterfact: Since tail entities are often not captured by models, and therefore are not suitable for testing modification edits, RippleEdit collects triplets about popular entities, where the subject corresponds to one of the top-viewed pages in Wikipedia.
  • Convsent: This is a sentiment editing task that assesses the model's ability to modify a dialog agent's sentiment on a specific topic without affecting its responses to other topics.
  • Sanitation: This dataset specifically addresses privacy concerns associated with learned language models.
dataset structure 
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
├── convsent
│   ├── trivia_qa_test.json
│   └── trivia_qa_train.json
└── wiki_recent
    ├── recent_test.json
    └── recent_train.json

Datasets for Chinese Knowledge: CKnowEdit

dataset HuggingFace WiseModel ModelScope Description
CKnowEdit [HuggingFace] [WiseModel] [ModelScope] dataset for editing Chinese Knowledge
  • Here, you can follow CKnowEdit.md to find more details about CKnowEdit and run Chinese knowledge editing experiments.
dataset description

CKnowEdit is a high-quality Chinese-language dataset for knowledge editing which is highly characterized by the Chinese language, with all data sourced from Chinese knowledge bases. It is meticulously designed to more deeply discern the nuances and challenges inherent in the comprehension of the Chinese language by current LLMs, providing a robust resource for refining Chinese-specific knowledge within LLMs.

The field descriptions for the data in CKnowEdit are as follows:

"prompt": query inputed to the model (str)
"target_old": the incorrect response previously generated by the model (str)
"target_new": the accurate answer of the prompt (str)
"portability_prompt": new prompts related to the target knowledge (list or None)
"portability_answer": accurate answers corresponding to the portability_prompt (list or None)
"locality_prompt": new prompts unrelated to the target knowledge (list or None)
"locality_answer": accurate answers corresponding to the locality_prompt (list or None)
"rephrase": alternative ways to phrase the original prompt (list)
dataset structure 
CknowEdit
├── Chinese Literary Knowledge
│   ├── Ancient Poetry
│   ├── Proverbs
│   └── Idioms
├── Chinese Linguistic Knowledge
│   ├── Phonetic Notation
│   └── Classical Chinese
├── Chinese Geographical Knowledge
└── Ruozhiba

Datasets for Factual Knowledge

dataset Google Drive BaiduNetDisk Description
ZsRE plus [Google Drive] [BaiduNetDisk] Question Answering dataset using question rephrasings
Counterfact plus [Google Drive] [BaiduNetDisk] Counterfact dataset using Entity replacement

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).
dataset description 
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
  • counterfact: original counterfact dataset using Entity replacement
  • zsre: original question answering dataset using question rephrasings
  • locality (evaluation for locality, see details in this paper)
    • Commonsense Task: evaluation for other downstream tasks such as commonsense task
    • Distracting Neighbor: test on distracting neighborhood (reference: Detecting Edit Failures...)
    • Other Attribution
  • portability
    • Inverse Relation: evaluation for one-to-one relationship such as spouse
    • One Hop: evaluation for one-hop reasoning
    • Subject Replace: evaluation for synonym replacement

Datasets for Conceptual Knowledge: ConceptEdit

dataset Google Drive HuggingFace Dataset Description
ConceptEdit [Google Drive] [HuggingFace Dataset] dataset for editing conceptual knowledge
  • Here, you can follow ConceptEdit.md to run concept editing experiments.
dataset description 
data
└──concept_data.json
    ├──final_gpt2_inter.json
    ├──final_gpt2_intra.json
    ├──final_gptj_inter.json
    ├──final_gptj_intra.json
    ├──final_llama2chat_inter.json
    ├──final_llama2chat_intra.json
    ├──final_mistral_inter.json
    └──final_mistral_intra.json

Concept Specific Evaluation Metrics

  • Instance Change: capturing the intricacies of these instance-level changes
  • Concept Consistency: the semantic similarity of generated concept definition

Datasets for Multimodal Knowledge: MMEdit

dataset Google Drive BaiduNetDisk Description
E-IC [Google Drive] [BaiduNetDisk] dataset for editing Image Captioning
E-VQA [Google Drive] [BaiduNetDisk] dataset for editing Visual Question Answering
  • All images used in E-IC and E-VQA are available for download at Google Drive
  • 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.
dataset description 
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
  • Multimodal locality (evaluation for multimodal locality, see dataset's details in this paper)

Datasets for detoxifying LLMs: SafeEdit

dataset HuggingFace Dataset Description
SafeEdit [HuggingFace Dataset] dataset for detoxifying LLMs
  • Here, you can follow SafeEdit.md to run detoxification editing experiments.
dataset description 
data
└──SafeEdit_train.json
└──SafeEdit_val.json
└──SafeEdit_test.json

Detoxifying Specific Evaluation Metrics

  • Defense Duccess (DS): the detoxification success rate of edited LLM for adversarial input (attack prompt + harmful question), which is used to modify LLM.
  • Defense Generalization (DG): the detoxification success rate of edited LLM for out-of-domain malicious inputs.
  • General Performance: the side effects for unrelated task performance.

Tutorial notebook

Method Description GPT-2 LlaMA
IKE In-Context Learning (ICL) Edit [Colab-gpt2] [Colab-llama]
ROME Locate-Then-Edit Neurons [Colab-gpt2] [Colab-llama]
MEMIT Locate-Then-Edit Neurons [Colab-gpt2] [Colab-llama]

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

Editing GPU memory usage

Our results are all based on the default configuration

llama-2-7B chatglm2 gpt-j-6b gpt-xl
FT 60GB 58GB 55GB 7GB
SERAC 42GB 32GB 31GB 10GB
IKE 52GB 38GB 38GB 10GB
MEND 46GB 37GB 37GB 13GB
KN 42GB 39GB 40GB 12GB
ROME 31GB 29GB 27GB 10GB
MEMIT 33GB 31GB 31GB 11GB
AdaLoRA 29GB 24GB 25GB 8GB
GRACE 27GB 23GB 6GB
WISE 34GB 27GB 7GB

📌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

## 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. [sequential_edit=True for continuous editing]

metrics, edited_model, _ = editor.edit(
    prompts=prompts,
    ground_truth=ground_truth,
    target_new=target_new,
    locality_inputs=locality_inputs,
    sequential_edit=False # True: start continuous editing ✈️
)
## metrics: edit success, rephrase success, locality e.g.
## edited_model: post-edit model

The maximum input length for EasyEdit is 512. If this length is exceeded, you will encounter the error "CUDA error: device-side assert triggered." You can modify the maximum length in the following file:LINK

Step7: RollBack In sequential editing, if you are not satisfied with the outcome of one of your edits and you do not wish to lose your previous edits, you can use the rollback feature to undo your previous edit. Currently, we only support the GRACE method. All you need to do is a single line of code, using the edit_key to revert your edit.

editor.rolllback('edit_key')

In EasyEdit, we default to using target_new as the edit_key

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()

Use EasyEdit with KnowEdit

Dataset

KnowEdit is a benchmark dataset of knowledge editing for LLMs. You can easily obtain KnowEdit from HuggingFace, HuggingFace, and ModelScope.

dataset HuggingFace HuggingFace ModelScope
KnowEdit [HuggingFace] [WiseModel] [ModelScope]

Usage

We provide detailed scripts for user to easily use KnowEdit, please refer to examples.

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)

Reliability Generalization Locality Portability
FT 56.94 52.02 96.32 51.03
SERAC 99.49 99.13 100.00 57.82
IKE 100.00 99.98 69.19 67.56
MEND 94.24 90.27 97.04 56.95
KN 28.95 28.43 65.43 37.18
ROME 92.45 87.04 99.63 57.47
MEMIT 92.94 85.97 99.49 60.64

We also present editing results of KnowEdit on LlaMA-2-7B using EasyEdit.

DataSet Metric SERAC ICE AdaLoRA MEND ROME MEMIT FT-L FT
WikiData_recent
Edit Succ. 98.68 60.74 65.61 76.88 85.08 85.32 71.18 31.24
Portability 63.52 36.93 47.22 50.11 37.45 37.94 48.71 15.91
Locality 100.00 33.34 55.78 92.87 66.2 64.78 63.7 3.65
Fluency 553.19 531.01 537.51 586.34 574.28 566.66 549.35 428.67
ZsRE
Edit Succ. 99.67 66.01 69.86 96.74 96.57 83.07 54.65 36.88
Portability 56.48 63.94 52.95 60.41 52.20 51.43 45.02 8.72
Locality 30.23 23.14 72.21 92.79 27.14 25.46 71.12 0.31
Fluency 410.89 541.14 532.82 524.33 570.47 559.72 474.18 471.29
WikiBio
Edit Succ. 99.69 95.53 97.02 93.66 95.05 94.29 66.27 95.64
Locality 69.79 47.90 57.87 69.51 46.96 51.56 60.14 13.38
Fluency 606.95 632.92 615.86 609.39 617.25 616.65 604.00 589.22
WikiData_counterfact
Edit Succ. 99.99 69.83 72.14 78.82 83.21 83.41 51.12 26.78
Portability 76.07 45.32 55.17 57.53 38.69 40.09 39.07 16.94
Locality 98.96 32.38 66.78 94.16 65.4 63.68 62.51 0.29
Fluency 549.91 547.22 553.85 588.94 578.84 568.58 544.80 483.71
ConvSent
Edit Succ. 62.75 52.78 44.89 50.76 45.79 44.75 49.50 61.93
Locality 0.26 49.73 0.18 3.42 0.00 0.00 0.00 0.00
Fluency 458.21 621.45 606.42 379.43 606.32 602.62 607.86 546.24
Sanitation
Edit Succ. 0.00 72.50 2.50 0.00 85.00 48.75 0.00 60.00
Locality 100.00 56.58 65.50 5.29 50.31 67.47 14.78 42.61
Fluency 416.29 794.15 330.44 407.18 465.12 466.10 439.10 351.39

❗️❗️ Please note that if you wish to reproduce the results regarding Rome on Knowedi, ensure that fp16: False.

For the locality metric, we calculate the score based on the proportion of tokens that remain unchanged before and after editing. For example, if the output tokens before editing are [29, 234, 334] and after editing are [29, 234, 333], the locality score for this data would be 66.67. For the portability metric, we calculate it by taking the average of all sub-scores under the portability category.

TO DO In next version, we plan to:
  • Explore and integrate more robust editing methods, focusing on locality and portability metrics.
  • Provide a comprehensive evaluation suite for editing methods, including fact modification, fact erasure and hallucination erasure.
  • Provide a causal analysis component for analyzing knowledge storage mechanisms.
  • knowledge editing for other tasks(except factual editing), like personality editing, etc.

Meanwhile, we will offer long-term maintenance to fix bugs, solve issues and meet new requests. So if you have any problems, please put issues to us.

Citation

Please cite our paper if you use EasyEdit 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}
}

@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}
}

@article{mao2023editing,
  title={Editing personality for llms},
  author={Mao, Shengyu and Zhang, Ningyu and Wang, Xiaohan and Wang, Mengru and Yao, Yunzhi and Jiang, Yong and Xie, Pengjun and Huang, Fei and Chen, Huajun},
  journal={arXiv preprint arXiv:2310.02168},
  year={2023}
}

@article{wang2024wise,
  title={WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models},
  author={Wang, Peng and Li, Zexi and Zhang, Ningyu and Xu, Ziwen and Yao, Yunzhi and Jiang, Yong and Xie, Pengjun and Huang, Fei and Chen, Huajun},
  journal={arXiv preprint arXiv:2405.14768},
  year={2024}
}

🎉Contributors

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

Other Related Projects

🙌 We would like to express our heartfelt gratitude for the contribution of FastEdit, ROME, GRACE, MELO, PMET to our project, as we have utilized portions of their source code in our project. Many thanks to all the colleagues in the community for submitting issues and providing technical support. Appreciation is also extended to all PR contributors, and issue feedback providers during the EasyEdit version iterations, especially ancelia06 for correcting the grammar of README.