This is the codebase for our paper "LongSafety: Evaluating Long-Context Safety of Large Language Models".

LongSafety is the first benchmark to comprehensively evaluate LLM safety in open-ended long-context tasks. It encompasses 1,543 instances with an average length of 5,424 words and comprises 7 safety issues and 6 task types, covering a wide range of long-context safety problems in real-world scenarios.

🎉 2025/05/16: Our Paper LongSafety: Evaluating Long-Context Safety of Large Language Models has been accepted by ACL 2025 Main Conference! Please feel free to contact us if you are interested in our work.

🎉 2025/02/26: We have released the data of LongSafety on Huggingface🤗.

🎉 2025/02/25: We have released the data and evaluation code of LongSafety.

The data of LongSafety is placed in data/longsafety_meta.jsonl and data/longsafety_docs.jsonl. The meta information of each instance is included in data/longsafety_meta.jsonl with the following format:

• id (integer): A unique indentifier of the instance.
• link (list): The web links of the source documents.
• length (integer): The word length of the long context.
• safety_type (string): The safety type of the instance.
• key_words (list): The safety keywords of the instance.
• instruction (string): The safety instruction.
• task_type (string): The task type of the instruction.
• doc_num (integer): The number of documents integrated in the long context.

The long context of each instance is included in data/longsafety_docs.jsonl. The format is as follows:

• id (integer): A unique indentifier of the instance.
• context (string): The actual long context.

The instances in the two files are associated by the key id.

To run the evaluation codes, please first install the necessary packages.

pip install -r requirements.txt

Then you can evaluate any desired models via eval.sh:

bash eval.sh

You can change the model_name and model_path in eval.sh to evaluate different models. You will also need to set your openai api key at OPENAI_API_KEY in eval.sh to use the multi-agent framework for safety judgment of the generation results. The results will be saved in ./result by default, and you can freely change it if necessary.

Please kindly cite our paper if you find our work helpful.

@misc{lu2025longsafetyevaluatinglongcontextsafety,
      title={LongSafety: Evaluating Long-Context Safety of Large Language Models}, 
      author={Yida Lu and Jiale Cheng and Zhexin Zhang and Shiyao Cui and Cunxiang Wang and Xiaotao Gu and Yuxiao Dong and Jie Tang and Hongning Wang and Minlie Huang},
      year={2025},
      eprint={2502.16971},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.16971}, 
}