| --- |
| license: mit |
| language: |
| - en |
| - zh |
| --- |
| ## Introduction |
| The ShieldLM model ([paper link](https://arxiv.org/abs/2402.16444)) 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](https://arxiv.org/abs/2402.16444) for more detailed evaluation results. |
