coding-agent-traces
收藏资源简介:
Coding Agent Traces数据集包含来自真实AI编码会话的每轮追踪数据,其中对话内容经过脱敏处理。该数据集通过Skein CLI工具收集,贡献者需明确同意并仅贡献自己的数据。数据以行为单位,每行代表一轮交互,包括助手响应、提示和请求设置。内容经过分层脱敏处理以移除机密信息和PII,但脱敏是尽力而为的,并非完全匿名化。数据字段涵盖schema、rowId、sessionKey、turnIndex、provider、model、prompt、output、reasoning、toolIO、outcome等。适用于研究代理编码中的真实提示和工具使用模式、跨提供者/模型/工具的令牌/成本/延迟分布基准测试、分析工具调用模式以及测量结果/错误率。使用注意事项包括不要尝试重新识别贡献者、假设可能残留敏感数据,并遵守源模型提供者的条款。数据集采用ODC-BY 1.0许可。
The Coding Agent Traces dataset contains per-turn trace data from real AI coding sessions, with conversation content anonymized. It is collected via the Skein CLI tool, with contributors explicitly consenting and contributing only their own data. Data is organized by rows, each representing a single interaction (including assistant responses, prompts, and request settings). Content undergoes hierarchical anonymization to remove confidential information and PII, though this is done on a best-effort basis and is not fully anonymous. Fields include schema, rowId, sessionKey, turnIndex, provider, model, prompt, output, reasoning, toolIO, outcome, and more. It is suitable for researching real-world prompt and tool usage patterns in agent coding, benchmarking token/cost/latency distributions across providers/models/tools, analyzing tool call patterns by model and effort level, and measuring outcome/error rates across configurations. Usage notes include not attempting to re-identify contributors, assuming residual personal/sensitive data may remain, and adhering to source model providers terms. The dataset is licensed under ODC-BY 1.0.




