Structured geological text dataset and source code for genetic type classification of Pb–Zn deposits

Abstract

This dataset accompanies the study “Natural language processing of structured geological texts for genetic type classification of Pb–Zn deposits.” It contains structured geological descriptions for 179 Pb–Zn deposits compiled from published geological literature. Each sample comprises four categories of geological information: mineral assemblage, wall-rock alteration, host-rock lithology, and ore textures and structures, together with a five-class deposit genetic-type label. Of the 179 samples, 139 were used for model development and 40 as an independent test set. Deposit names and explicit type-revealing statements were excluded from the model inputs to reduce label leakage. The repository also provides the source code used for data preprocessing, grouped stratified five-fold cross-validation, hyperparameter selection, model training, independent testing, ablation analysis, and prediction using the MacBERT-based hierarchical classification framework. Dataset fields, label definitions, software dependencies, and execution instructions are provided in the accompanying README file.