AIM-Intelligence/COMPASS-Policy-aware-SFT-Dataset

该数据集包含4,121个查询-响应对，用于COMPASS论文中的“留一域出”（LODO）政策感知监督微调（SFT）实验。数据集的目的是学习组织特定的政策边界（允许列表/拒绝列表），而不仅仅是通用的“安全/不安全”模式。响应是从在COMPASS评估中实现完全政策遵循的模型输出中选择的，并与相应的查询配对。数据集覆盖了7个领域：汽车（AutoViaMotors）、政府（CityGov）、金融（FinSecure）、医疗（MediCarePlus）、旅游（PlanMyTrip）、教育（TutoraVerse）和人力资源/招聘（VirtuRecruit），而电信领域（TelePath）被保留用于评估。数据集格式为聊天式，每个示例包含唯一的ID、公司名称、消息列表、查询类型、模型和来源。

This dataset contains 4,121 query–response pairs used for the Leave-One-Domain-Out (LODO) policy-aware supervised fine-tuning (SFT) experiment in the COMPASS paper. The purpose is policy-aware SFT for learning organization-specific policy boundaries (allowlist/denylist), beyond generic “safe/unsafe” patterns. Responses are selected from model outputs that achieved full policy adherence under COMPASS evaluation, paired with their corresponding queries. The dataset covers 7 domains: AutoViaMotors (Automotive), CityGov (Government), FinSecure (Financial), MediCarePlus (Healthcare), PlanMyTrip (Travel), TutoraVerse (Education), and VirtuRecruit (HR/Recruiting), while the TelePath (Telecom) domain is held out for evaluation. The data format is chat-based, with each example containing a unique ID, company name, messages list, query type, model, and source.