借鉴 Evol Instruct 方法，利用 Deepseek R1 推理模型的能力，合成可观测性和根因分析数据。 Leveraging the Evol Instruct method and the capabilities of Deepseek R1 inference model to synthesize observability and root cause analysis data.
-
根因分析问题的 seed 冷启动 Cold start seed for root cause analysis problems
-
根因分析问题的 seed 从业内公开故障复盘文档热启动 Hot start seed for root cause analysis problems from publicly available industry incident postmortems
-
考虑 self_instruct 方法，从 service_list 中生产更多问题，但要考虑是否会重复率过高(试了几次，10 个冷启动的质量过滤都不能 100% 通过) Consider self_instruct method to generate more questions from service_list, but need to consider if the duplication rate would be too high (tried several times, 10 cold starts couldn't pass 100% quality filtering)
-
补充问题的背景 Enrich context to the problems
-
扩展更多 Depth 和 Breath 问题 Evol instruct for more Depth and Breadth questions
-
问题质量检测过滤器 Question quality detection filter
-
生成问题对应的监控指标的行为模式 Generate behavior patterns of monitoring metrics corresponding to the problems
-
根据模式生成指标数据的 Python 代码 Generate Python code for metric data based on patterns
-
生成问题对应的日志和变更事件 Generate corresponding logs and change events for the problems
-
生成问题对应的调用链数据(触发敏感审查了，奇怪) Generate corresponding trace data for the problems (triggered sensitive review, strange)
-
生成 RCA 报告 Generate RCA reports
-
生成数据和报告之间的一致性检查 Generate consistency check between data and reports
-
单独脚本可运行 Individual scripts are runnable
-
主程序可运行 Main program is runnable
-
批量 incident/metric 要按 id 存储文件，方便多次运行 Batch incident/metric files should be stored by id for convenient multiple runs
-
考虑一致性检查的后续处理 Consider follow-up processing for consistency checks
-
考虑 metric/log data 超过 window 大小的方案 Consider solutions for metric data exceeding window size
感谢以下项目和工具的启发与帮助： Thanks to the following projects and tools for their inspiration and help:
-
distilabel: 提供了高质量数据标注和评估的初始思路 distilabel: Provided initial ideas for high-quality data labeling and evaluation
-
flip.ai: 提供了可观测性领域 evol-instruct 和 RLHAIF 的思路 flip.ai: Provided ideas for evol-instruct and RLHAIF in the observability domain
-
ChatTS: 为时序数据生成提供了指标属性和常见模式的样例 ChatTS: Provided examples of metric properties and common patterns for time series data generation
-
MCQ2: 为时序数据生成提供了 python 合成的思路 MCQ2: Provided ideas for Python synthesis of time series data
-
cursor: 提供了优秀的 AI 辅助编程体验 cursor: Provided excellent AI-assisted programming experience
-
deepseek: 提供了强大的开源推理模型能力，为大批量合成数据提供了可能性 deepseek: Provided powerful open-source inference model capabilities, making large-scale data synthesis possible