moralogyengine/moralogy-1200

moralogy-1200是一个基于公理的道德推理数据集，包含1,200个直接偏好优化（DPO）对。该数据集从Moralogy框架中确定性生成，无需人工标注或GPT-4调用，完全基于公理推导。数据集涵盖四个领域的伦理困境场景：医疗分诊、军事/国防AI、自动驾驶AI和客户服务AI。每条记录包含唯一标识符、领域特定系统提示、伦理困境场景、正确应用H/Consent/PGH的响应、未能通过谓词评估的响应、崩溃状态、领域标签、拒绝响应中的失败模式以及源场景包。数据集设计用于与Moral Kernel（一个确定性谓词评估层）结合使用，采用Apache 2.0许可证开放，是Moralogy Engine项目的一部分。

moralogy-1200 is an axiomatic moral reasoning dataset comprising 1,200 DPO (Direct Preference Optimization) pairs. Generated deterministically from the Moralogy framework without human annotation or GPT-4 calls, it is derived purely from axioms. The dataset features ethical dilemma scenarios across four domains: Medical Triage, Military/Defense AI, Autonomous Vehicles AI, and Customer Service AI. Each record includes fields such as a unique identifier, domain-specific system prompt, ethical dilemma scenario, response correctly applying H/Consent/PGH, response that fails predicate evaluation, collapse state, domain label, failure mode in the rejected response, and source scenario bundle. Designed for use with the Moral Kernel—a deterministic predicate evaluation layer—the dataset is open under the Apache 2.0 license and forms part of the Moralogy Engine project.