2026-10-13T07:48:03Z https://www.scidb.cn/oai

10.57760/sciencedb.j00133.00687 2026-06-29T08:56:39Z

2026-06-29 Experimental Dataset for A Lightweight Content Safety Detection Model via Pruning-Compression and Memory-Monitoring doi:10.57760/sciencedb.j00133.00687 en This dataset provides the supporting data for the paper “A Lightweight Content Safety Detection Model Based on Pruning Compression and Memory Monitoring.” It is mainly used to reproduce the experimental results and verify model performance. The dataset is derived from the public datasets ChineseHarm-Bench and ChineseSafe. Based on these datasets, model training and testing were conducted, and the experimental result data were compiled. The data generation process included constructing the content safety detection model LCSD-PCMM based on a pretrained language model, applying structured pruning compression to the model, and integrating a memory monitoring module. The model was then trained on the standard training set and used for inference on the test set to obtain various evaluation metrics.The data files are categorized as follows:(1) Experimental result data file (Excel format)This file supports the tables and figures in the paper. It contains two worksheets:The first worksheet corresponds to Tables 1, 2, and 3 in the paper, recording the overall model performance comparison, ablation experiment results, and pruning metric comparison results, respectively.The second worksheet corresponds to Figures 2, 3, and 4, recording performance changes under different pruning rates and model performance across different categories.(2) Baseline model test result data file (JSON format)This file records the prediction results of the baseline models on the test set. Each data entry contains the input text, true label, predicted label, and whether the prediction is correct.(3) Original data file (JSON format)This file contains the original text data used for model training and testing, divided into training, validation, and test sets. Each data entry consists of text content and its corresponding label.All result data are used to generate the tables and figures in the paper, while the original data are used for model training and evaluation.The data in this dataset mainly include two types: structured tabular data and text data.For the experimental result data in Excel, each table organizes data by row, with each row representing an experimental configuration or statistical result. The column elements vary according to the corresponding table or figure, and mainly include the following types of information: model name, experimental setting, dataset category, performance metrics, and inference time. The performance metrics include accuracy (Acc) and Macro-F1, both of which are dimensionless indicators. Inference time (T) is measured in seconds and represents the average time taken by the model during the testing stage.For category-level statistical data (corresponding to Figures 3 and 4), each row represents a specific category, such as non-violation or vulgar/pornographic content, and the column elements represent the F1 scores of different models for that category.For JSON-format data files:In the baseline model test result data, each record represents a test sample and contains the text content, true label, predicted label, and whether the prediction result is correct.In the original data file, each record represents a sample consisting of text content and its corresponding label, divided into the training set, validation set, and test set.The dataset contains no missing values, and all results are derived from actual experimental outputs. However, it should be noted that due to random initialization during deep learning model training and differences in hardware environments, the experimental results may show slight fluctuations, with errors typically kept within a reasonable range (generally no more than ±0.5%). In addition, because complete per-sample prediction outputs were not retained for some experiments, this dataset mainly provides aggregated evaluation metric data, but this does not affect the reproduction and verification of the paper’s conclusions. content safety; pruning compression; experimental dataset Nankai Lin zhang bo wei Yuying Li Shifa Dai Zhen Xu Lianxi Wang PUBLIC https://creativecommons.org/licenses/by-nc-sa/4.0/ dataset Science Data Bank