neo4j/aip-skillbench-24task-sonnet-aipv0_3a3
收藏资源简介:
AIP-SkillBench数据集是一个用于评估两种技能格式在相同任务上性能的原始评估运行数据。具体来说,它进行了头对头比较:一种是human-curated,即任务原始的人类编写技能(散文形式);另一种是aip-from-curated,即将相同的人类技能编译成AIP(一种模式验证的执行图表示)。数据集包含24个任务的分层样本,分为三个平衡的8任务队列(A、B、C)。每个任务×模式运行5次独立试验,总计240次运行。求解器代理为claude-agent-acp,模型为claude-sonnet-4-6,AIP编写使用claude-opus基于AIP规范v0.3a3完成。沙箱环境为docker,基准测试扩展自SkillsBench。数据布局包括每个队列的文件夹,包含campaign.json、status.json、summary.csv、summary.jsonl、cells和logs等文件。源代码和技能可在GitHub仓库中找到,标签为sonnet-aipv0.3a3。数据集可用于分析和比较不同技能格式在LLM代理任务中的效果。
The AIP-SkillBench dataset is a raw evaluation run dataset designed to assess the performance of two skill formats on identical tasks. Specifically, it implements a head-to-head comparison: one format consists of human-curated, original human-written task skills presented in prose form; the other is aip-from-curated, which compiles the same human-written skills into AIP, a pattern-validated execution graph representation. The dataset comprises stratified samples of 24 tasks, split into three balanced 8-task queues labeled A, B, and C. Each task×mode combination undergoes 5 independent trials, resulting in a total of 240 runs. The solver agent used is claude-agent-acp, with the underlying model being claude-sonnet-4-6. AIP authoring was completed using claude-opus based on the AIP Specification v0.3a3. The sandbox environment is Docker, and the benchmark is extended from SkillsBench. The data layout includes folders for each queue, containing files such as campaign.json, status.json, summary.csv, summary.jsonl, cells, and logs. The source code and skills are available in a GitHub repository tagged with sonnet-aipv0.3a3. This dataset can be used to analyze and compare the performance of different skill formats in LLM agent task scenarios.




