README.md

metadata

datasets:
  - zjunlp/ChineseHarm-bench
language:
  - zh
license: cc-by-nc-4.0
pipeline_tag: text-classification
library_name: transformers

🚀Installation

1. Clone the repositories:

   git clone https://github.com/zjunlp/ChineseHarm-bench
   cd ChineseHarm-bench
   git clone https://github.com/hiyouga/LLaMA-Factory
   

2. Install dependencies:

   cd LLaMA-Factory
   pip install -e ".[torch,metrics]" 
   

📚Inference

We release the following variants of our harmful content detection model:

• ChineseGuard-1.5B
• ChineseGuard-3B
• ChineseGuard-7B

🔹 Single Inference (Example)

Run single-input inference using the ChineseGuard-1.5B model:

SCRIPT_PATH="../infer/single_infer.py"
model_name="zjunlp/ChineseGuard-1.5B"
text="代发短信，有想做的联系我，无押金"

python $SCRIPT_PATH \
    --model_name $model_name \
    --text $text

🔸 Batch Inference (Multi-NPU or Multi-GPU)

To run inference on the entire ChineseHarm-Bench using ChineseGuard-1.5B and 8 NPUs:

SCRIPT_PATH="../infer/batch_infer.py"
model_name="zjunlp/ChineseHarm-1.5B"
file_name="../benchmark/bench.json"
output_file="../benchmark/bench_ChineseHarm-1.5B.json"

python $SCRIPT_PATH \
    --model_name $model_name \
    --file_name $file_name \
    --output_file $output_file \
    --num_npus 8

For more configuration options (e.g., batch size, device selection, custom prompt templates), please refer to single_infer.py and batch_infer.py.

Note: The inference scripts support both NPU and GPU devices.

Evaluation: Calculating F1 Score

After inference, evaluate the predictions by computing the F1 score with the following command:

python ../calculate_metrics.py \
    --file_path "../benchmark/bench_ChineseHarm-1.5B.json" \
    --true_label_field "标签" \
    --predicted_label_field "predict_label"

🚩Citation

Please cite our repository if you use ChineseHarm-bench in your work. Thanks!

@misc{liu2025chineseharmbenchchineseharmfulcontent,
      title={ChineseHarm-Bench: A Chinese Harmful Content Detection Benchmark}, 
      author={Kangwei Liu and Siyuan Cheng and Bozhong Tian and Xiaozhuan Liang and Yuyang Yin and Meng Han and Ningyu Zhang and Bryan Hooi and Xi Chen and Shumin Deng},
      year={2025},
      eprint={2506.10960},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.10960}, 
}

Codebase: https://github.com/zjunlp/ChineseHarm-bench