domyn/DomynGuardFinancialBenchmark

Domyn Guard金融安全基准数据集是一个用于评估金融领域AI安全分类器性能的基准测试集。该数据集包含106,307个来自FinSafeGuard金融安全测试集的对话样本，每个样本都提供了标准真实标签（包括安全标签、风险类别和严重性等级）与Domyn Guard模型的预测结果（预测标签、风险类别、严重性等级以及完整的推理轨迹）的对比。数据集覆盖银行、金融和保险三个金融领域，包含用户输入、多轮对话和包含助手回复的对话三种对话类型。数据集采用20个细粒度风险类别（分为5个超类别）和4个严重性等级的分类体系。主要用途包括：检查Domyn Guard模型在大型BFSI安全测试集上的行为、比较不同安全护栏模型在相同提示和真实标签下的性能、按金融领域、对话类型、风险类别和严重性等级进行错误分析。

Domyn Guard Financial Safety Benchmark is a benchmark dataset for evaluating the performance of AI safety classifiers in the financial domain. It contains 106,307 conversation samples from the FinSafeGuard financial-safety test set, each providing a comparison between canonical ground-truth labels (including safety label, risk category, and severity level) and Domyn Guard model predictions (predicted label, category, severity, and full reasoning trace). The dataset covers three financial domains: banking, financial, and insurance, and includes three conversation types: user input, multi-turn conversations, and conversations with assistant replies. It employs a classification system with 20 fine-grained risk categories (organized into 5 super-categories) and 4 severity levels. Primary uses include: inspecting Domyn Guards behavior on a large BFSI safety test set, comparing different guardrail models against the same prompts and ground truth, and analyzing errors by slicing across financial domains, conversation types, risk categories, and severity levels.