lincyaw/openrca2-lite-v3

这是一个用于根因分析（RCA）评估的数据集，包含500个精选案例，重建于2026年5月3日，使用manifest驱动的因果图推理器生成。数据集基于1464个原始案例池（与v1/v2版本相同），但每个案例的真实因果图（causal_graph.json）都通过新的故障manifest扩展重新生成，替代了旧有的仅基于规则的传播器。与v2相比，新推理器产生的图结构更均匀，平均度数更小，幻觉传输节点更少，并且有41个案例在新推理器下出现循环服务图。数据集案例来源主要为AegisLab的detector_success（包括320个ts、142个hs和38个otel-demo系统），覆盖多种混沌家族（如hybrid_clean、Network*、Pod*等），并按最长路径、服务数量等指标进行统计和筛选。数据布局包括manifest.jsonl文件和各案例目录，包含因果图、注入配置、环境信息等文件。A 500-case curated RCA evaluation dataset, rebuilt on 2026-05-03 with the manifest-driven causal-graph reasoner. It uses the same 1464-case raw pool as v1/v2, but the ground-truth causal_graph.json for every case is regenerated by the new fault-kind manifest expansion instead of the legacy rule-only propagator. Compared to v2, the new reasoner produces more uniform graph structures with smaller mean degree, fewer hallucinated transit nodes, and 41 cases now have cyclic service graphs. The dataset consists of 500 cases sourced primarily from AegisLab detector_success (320 ts, 142 hs, 38 otel-demo), covering various chaos families (e.g., hybrid_clean, Network*, Pod*), and is organized with statistics on longest path, number of services, etc. The layout includes a manifest.jsonl file and case directories containing causal graphs, injection configurations, environment data, and more.