neo4j/aip-skillbench-3med-sonnet-aipv0_3a2
收藏资源简介:
AIP-SkillBench数据集包含3个中等难度任务的评估运行数据,这些任务是从SkillsBench中随机抽样的。该数据集评估了三种技能格式在同一任务上的表现:human-curated(人类编写的原始技能)、aip-from-curated(将人类技能编译为AIP格式,即经过模式验证的执行图表示)和aip-from-instruction(仅从任务指令编写的AIP技能,无人类技能参与)。每个任务×模式组合进行了5次独立试验,总共有45次运行。使用的求解代理是claude-agent-acp,模型是claude-sonnet-4-6,AIP编写基于AIP规范v0.3a2,使用claude-opus模型完成,沙箱环境为docker。数据集扩展自SkillsBench基准,任务包括晶体学Wyckoff位置分析、无人机规划控制和地震板块计算。数据布局包括campaign.json(运行矩阵)、status.json(运行统计)、summary.csv(每试验一行数据)以及cells/和logs/目录(分别存储每试验的工作目录和求解日志)。summary.csv包含任务、模型、模式、试验、状态、奖励、工具调用次数、运行时间、错误等列。
The AIP-SkillBench dataset contains evaluation run data for three moderate-difficulty tasks randomly sampled from the SkillsBench benchmark. This dataset evaluates the performance of three skill formats on the same tasks: human-curated (original human-written skills), aip-from-curated (human skills compiled into AIP format, i.e., schema-validated execution graph representations), and aip-from-instruction (AIP skills written solely from task instructions without human skill involvement). Each task × mode combination underwent 5 independent trials, resulting in a total of 45 runs. The solving agent used is claude-agent-acp, with the underlying model being claude-sonnet-4-6. AIP scripts are developed based on AIP Specification v0.3a2, and the generation process was completed using the claude-opus model. The sandbox environment utilized is Docker. This dataset is extended from the SkillsBench benchmark, and the included tasks cover crystallographic Wyckoff position analysis, drone planning and control, and seismic plate calculation. The dataset's structure includes campaign.json (run matrix), status.json (run statistics), summary.csv (one row per trial), as well as the cells/ and logs/ directories, which respectively store the working directory and solution logs for each trial. The summary.csv file contains columns including task, model, mode, trial, status, reward, number of tool calls, runtime, and errors.




