ArtificialAnalysis/AA-Briefcase-Lite
收藏资源简介:
AA-Briefcase-Lite是AA-Briefcase的公共示例场景,由Artificial Analysis开发,用于前沿AI代理在现实、长期知识工作方面的评估。该数据集扩展了前沿模型基准测试的范围,超越了编码和短形式推理,涵盖了知识工作者日常产生的专业交付成果。数据集包含一个商业尽职调查场景,模拟美国中型市场私募股权公司Halberd Capital Partners对新西兰鸡蛋生产商Aurora Eggs Ltd的收购前尽职调查。场景基于新西兰鸡蛋市场的真实数据,包含真实和合成文档,以及需要代理识别和协调的跨来源矛盾。数据集包括一周的任务,涵盖四个可独立完成的交付成果:市场结构概述(LaTeX/PDF)、市场规模和预测模型(Excel)、目标评估(PowerPoint)以及初步发现简报视频(MP4/SRT)。源文件池包含67个来源(147个文件),分为共享文件和每周文件。数据集文件包括checks.jsonl(63个检查项)和tasks.jsonl(4个任务)。评估方法基于Stirrup框架,在离线沙箱中进行,使用前沿模型作为评委进行分级。数据集还包含用于生成和分级的提示模板,以及六个前沿模型的示例提交。
AA-Briefcase-Lite is the public example scenario for AA-Briefcase, Artificial Analysis frontier agentic evaluation of realistic, long-horizon knowledge work. It extends frontier model benchmarking beyond coding and short-form reasoning to the professional deliverables knowledge workers produce day to day. The dataset features a commercial due-diligence engagement where the evaluated model assists a Vice President at Halberd Capital Partners, a fictional US mid-market private-equity firm, in conducting outside-in commercial due diligence on Aurora Eggs Ltd, a privately-held New Zealand egg producer. The scenario is grounded in real data from the New Zealand egg market and uses real and synthetic documents, including deliberate cross-source contradictions. It contains one week of tasks with four independently completable deliverables: market overview (LaTeX/PDF), market sizing and forecast model (Excel), target assessment (PowerPoint), and preliminary findings briefing video (MP4/SRT). The source pool holds 67 sources (147 files) in shared and week-specific groups. Dataset files include checks.jsonl (63 checks) and tasks.jsonl (4 tasks). The methodology is built on the Stirrup framework, using an offline sandbox environment and graded by a panel of frontier models. The dataset also includes prompts for generation and binary rubric grading, as well as example submissions from six frontier models.




