Jazhyc/wildguardmix-annotations

该数据集包含了来自allenai/wildguardmix训练子集的90%的数据，并增加了经过微调的ModernBERT-large分类器组合的预测结果。数据集被划分为标注集，包含77,646个样本，并根据prompt_harm_label、adversarial和subcategory进行了分层。新增了预测有害概率、预测标签和真实标签等列。该数据集的集成分类器在单独的测试集上达到了95.8%的F1分数。该数据集旨在用于标注具有自然语言描述的意图提示的困难示例。

This dataset contains a 90% split of the allenai/wildguardmix train subset, augmented with predictions from a fine-tuned ModernBERT-large classifier ensemble. It is split into an annotation set with 77,646 samples, stratified on prompt_harm_label, adversarial, and subcategory. The dataset includes added columns for harmful probability, predicted label, and ground truth label. The ensemble classifier achieved an F1 score of 95.8% on a separate test set. The dataset is designed for annotating difficult examples with natural language descriptions of prompt intent.