A Benchmark for Semantic Sensitive Information in LLMs Outputs
Authors: Qingjie Zhang, Han Qiu, Di Wang, Yiming Li, Tianwei Zhang, Wenyu Zhu, Haiqin Weng, Liu Yan, Chao Zhang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | First, we construct a comprehensive and labeled dataset of semantic sensitive information, Sem SI-Set, by including three typical categories of Sem SI. Then, we propose a large-scale benchmark, Sem SI-Bench, to systematically evaluate semantic sensitive information in 25 SOTA LLMs. Our finding reveals that Sem SI widely exists in SOTA LLMs outputs by querying with simple natural questions. We open-source our project at https://semsi-project.github.io/. |
| Researcher Affiliation | Collaboration | Qingjie Zhang1, Han Qiu1 , Di Wang1, Yiming Li2, Tianwei Zhang2, Wenyu Zhu3, Haiqing Weng4, Liu Yan4, and Chao Zhang1 1Tsinghua University, 2Nanyang Technological University, 3Ascend Grace Tech, 4Ant Group Emails:{qj-zhang24@mails., qiuhan@}tsinghua.edu.cn |
| Pseudocode | No | The paper describes methods and processes verbally and visually (e.g., Figure 2 pipeline overview), but does not contain any explicit pseudocode blocks or algorithms. |
| Open Source Code | Yes | We open-source our project at https://semsi-project.github.io/. The codes for reproducing our results are provided in our project website: https://semsi-project.github.io/. |
| Open Datasets | Yes | First, we construct a comprehensive and labeled dataset of semantic sensitive information, Sem SI-Set, by including three typical categories of Sem SI. ... We open-source our project at https://semsi-project.github.io/. |
| Dataset Splits | Yes | We compress Sem SI-Set to a coreset of 1,000 samples, Sem SI-c Set, for labeling and benchmarking. ... We observe that if we proportionally reduce the occurrence of Sem SI of one model (e.g. GPT3.5-Turbo in Table 3), its metrics are almost the same after compression. What s more, the metrics are also the same for other models (e.g. GPT4o, Llama3-8B, and GLM4-9B in Table 3). We can see that the diffrence of metric values between the compressed and the original dataset is very close to 0. This implies a common coreset Sem SI-c Set, to represent Sem SI-Set and efficiently make Sem SI-Bench. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions the general timeframe of experiments: "Experiments of GPT-o1 series are done at the end of September 2024 while other experiments are done at August 2024." |
| Software Dependencies | No | The paper mentions accessing LLMs via public API and Hugging Face, and using GPT-4o for labeling, but it does not specify any particular software libraries, frameworks, or their version numbers that were used to implement their own methodology or analysis. |
| Experiment Setup | No | The paper focuses on benchmarking existing LLMs and describes the process of prompt generation, labeling, and metric computation. It does not involve training its own models, and therefore, does not provide experimental setup details like hyperparameters, optimizers, or training schedules which are typically associated with training a model. |