Datasets:
ArXiv:
License:
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,84 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-sa-4.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-nc-sa-4.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
<div align="center">
|
| 7 |
+
|
| 8 |
+
**Editing Conceptual Knowledge for Large Language Models**
|
| 9 |
+
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
<p align="center">
|
| 13 |
+
<a href="#-conceptual-knowledge-editing">Overview</a> •
|
| 14 |
+
<a href="#-usage">How To Use</a> •
|
| 15 |
+
<a href="#-citation">Citation</a> •
|
| 16 |
+
<a href="https://arxiv.org/abs/2403.06259">Paper</a> •
|
| 17 |
+
<a href="https://zjunlp.github.io/project/ConceptEdit">Website</a>
|
| 18 |
+
</p>
|
| 19 |
+
</div>
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
## 💡 Conceptual Knowledge Editing
|
| 23 |
+
|
| 24 |
+
<div align=center>
|
| 25 |
+
<img src="./flow1.gif" width="70%" height="70%" />
|
| 26 |
+
</div>
|
| 27 |
+
|
| 28 |
+
### Task Definition
|
| 29 |
+
|
| 30 |
+
**Concept** is a generalization of the world in the process of cognition, which represents the shared features and essential characteristics of a class of entities.
|
| 31 |
+
Therefore, the endeavor of concept editing aims to modify the definition of concepts, thereby altering the behavior of LLMs when processing these concepts.
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
### Evaluation
|
| 35 |
+
|
| 36 |
+
To analyze conceptual knowledge modification, we adopt the metrics for factual editing (the target is the concept $C$ rather than factual instance $t$).
|
| 37 |
+
|
| 38 |
+
- `Reliability`: the success rate of editing with a given editing description
|
| 39 |
+
- `Generalization`: the success rate of editing **within** the editing scope
|
| 40 |
+
- `Locality`: whether the model's output changes after editing for unrelated inputs
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
Concept Specific Evaluation Metrics
|
| 44 |
+
|
| 45 |
+
- `Instance Change`: capturing the intricacies of these instance-level changes
|
| 46 |
+
- `Concept Consistency`: the semantic similarity of generated concept definition
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
## 🌟 Usage
|
| 50 |
+
|
| 51 |
+
### 🎍 Current Implementation
|
| 52 |
+
As the main Table of our paper, four editing methods are supported for conceptual knowledge editing.
|
| 53 |
+
| **Method** | GPT-2 | GPT-J | LlaMA2-13B-Chat | Mistral-7B-v0.1
|
| 54 |
+
| :--------------: | :--------------: | :--------------: | :--------------: | :--------------: |
|
| 55 |
+
| FT | ✅ | ✅ | ✅ | ✅ |
|
| 56 |
+
| ROME | ✅ | ✅ |✅ | ✅ |
|
| 57 |
+
| MEMIT | ✅ | ✅ | ✅| ✅ |
|
| 58 |
+
| PROMPT | ✅ | ✅ | ✅ | ✅ |
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
### 💻 Run
|
| 62 |
+
You can follow [EasyEdit](https://github.com/zjunlp/EasyEdit/edit/main/examples/ConceptEdit.md) to run the experiments.
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
## 📖 Citation
|
| 66 |
+
|
| 67 |
+
Please cite our paper if you use **ConceptEdit** in your work.
|
| 68 |
+
|
| 69 |
+
```bibtex
|
| 70 |
+
@misc{wang2024editing,
|
| 71 |
+
title={Editing Conceptual Knowledge for Large Language Models},
|
| 72 |
+
author={Xiaohan Wang and Shengyu Mao and Ningyu Zhang and Shumin Deng and Yunzhi Yao and Yue Shen and Lei Liang and Jinjie Gu and Huajun Chen},
|
| 73 |
+
year={2024},
|
| 74 |
+
eprint={2403.06259},
|
| 75 |
+
archivePrefix={arXiv},
|
| 76 |
+
primaryClass={cs.CL}
|
| 77 |
+
}
|
| 78 |
+
```
|
| 79 |
+
|
| 80 |
+
## 🎉 Acknowledgement
|
| 81 |
+
|
| 82 |
+
We would like to express our sincere gratitude to [DBpedia](https://www.dbpedia.org/resources/ontology/),[Wikidata](https://www.wikidata.org/wiki/Wikidata:Introduction),[OntoProbe-PLMs](https://github.com/vickywu1022/OntoProbe-PLMs) and [ROME](https://github.com/kmeng01/rome).
|
| 83 |
+
|
| 84 |
+
Their contributions are invaluable to the advancement of our work.
|