domyn/FinSafeGuard
收藏资源简介:
FinSafeGuard是一个以质量为导向的合成数据集,专为金融AI安全设计,用于训练和评估银行、金融服务和保险(BFSI)对话中的安全分类器和护栏模型。该数据集通过FinSafeGuard流程生成,该流程通过四个阶段(预生成研究、条件自动进化、两层去重和LLM-as-Judge过滤)将2640万原始合成样本转化为1430万高置信度样本。此版本为Ultra-Mini变体,包含709,303个安全标注的对话,涵盖20个BFSI风险类别,适用于监督微调(SFT)安全分类器和基于推理的安全判断。
FinSafeGuard is a quality-driven synthetic dataset for financial AI safety, designed for training and evaluating safety classifiers and guardrail models on Banking, Financial Services, and Insurance (BFSI) conversations. The dataset is produced by the FinSafeGuard pipeline, which transforms 26.4M raw synthetic samples into 14.3M high-confidence examples through a four-stage process: pre-generation research, conditional auto-evolution, two-layer deduplication, and LLM-as-Judge filtering. This release is the Ultra-Mini variant — 709,303 safety-annotated conversations spanning 20 BFSI risk categories in instruction-tuning format, ready for supervised fine-tuning (SFT) of safety classifiers and reasoning-based safety judges.




