Outer-Void/blux-ca-dataset

BLUX-cA数据集是一个统一的提示和评估探针数据集，用于BLUX-cA宪法核心。该数据集与模型无关，提供了一个可直接发布的数据集（无训练流程），包含验证工具、发布门控和双重许可。数据集涵盖多个领域，每个领域包含500个示例，如核心（身份、伦理、边界设定、操纵检测）、安全（拒绝、重定向、安全框架）、治理（权力、机构、问责制）、编码（调试纪律、安全模式、拒绝利用请求）、推理（结构化思维、假设检查、权衡）、创建（提案、计划、文档、结构化输出）、对话（简洁、基于事实的对话）、效率（压缩、摘要、最小化问题、清晰度）和关系（边界、冲突降级、问责制、反操纵）。数据集采用JSONL格式，每条记录包含系统提示、用户内容和助手内容，强调审计笔记和确定性资产。

The BLUX-cA Dataset is a unified collection of prompts and evaluation probes for the BLUX-cA constitutional spine. It is model-agnostic and provides a ready-to-publish dataset (without a training pipeline) along with validation tooling, release gating, and dual licensing. The dataset spans multiple domains, each containing 500 examples, such as core (identity, ethics, boundary-setting, manipulation detection), safety (refusals, redirection, safety framing), governance (power, institutions, accountability), coding (debugging discipline, secure patterns, refusing exploit requests), reasoning (structured thinking, assumption checks, tradeoffs), creation (proposals, plans, documents, structured outputs), conversation (concise, grounded dialogue), efficiency (compression, bullet summaries, minimal questions, clarity), and relationships (boundaries, conflict de-escalation, accountability, anti-manipulation). The dataset uses JSONL format, with each entry containing a system prompt, user content, and assistant content, emphasizing Audit Notes and deterministic assets.