File size: 920 Bytes
b53ace0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
---
license: mit
language:
- en
- zh
---
## Introduction
The ShieldLM model ([paper link](xxx)) initialized from [Qwen-14B-Chat](https://huggingface.co/Qwen/Qwen-14B-Chat). ShieldLM is a bilingual (Chinese and English) safety detector that mainly aims to help to detect safety issues in LLMs' generations. It aligns with general human safety standards, supports fine-grained customizable detection rules, and provides explanations for its decisions. 
Refer to our [github repository](https://github.com/thu-coai/ShieldLM) for more detailed information.

## Usage
Please refer to our [github repository](https://github.com/thu-coai/ShieldLM) for the detailed usage instructions.

## Performance
ShieldLM demonstrates impressive detection performance across 4 ID and OOD test sets, compared to strong baselines such as GPT-4, Llama Guard and Perspective API.
Refer to [our paper](xxx) for more detailed evaluation results.