cyberforge-gpt56-luna-teacher-trajectories
收藏资源简介:
CyberForge — GPT-5.6-Luna Teacher Trajectories是一个专门用于学生模型监督微调的智能体轨迹数据集。该数据集包含通过Azure Responses API使用GPT-5.6-Luna模型在SEC-bench/Vuljector漏洞任务上进行推理求解时产生的多轮工具使用轨迹,完整保留了教师的推理过程。当前版本包含1,115条验证成功的轨迹,按漏洞来源和任务模式分为:fuzz-patch(592条)、fuzz-poc-san(225条)、workflow-patch(268条)、workflow-poc-san(30条)。数据集中1,091条轨迹包含至少一轮推理摘要,每轮推理覆盖率约为35%。每条数据采用结构化格式,包括多轮对话消息(如系统提示、用户输入、助手响应含推理内容和工具调用、工具返回结果)以及详细的元数据(如项目名称、漏洞ID、任务模式、漏洞来源、教师模型、教师机制、生成方法、对话轮次)。数据集已进行安全脱敏处理,确保不包含真实的API密钥或令牌。该数据集是正在进行中的部分导出快照,完整运行计划包含更多数据量。
CyberForge — GPT-5.6-Luna Teacher Trajectories is an intelligent agent trajectory dataset specifically designed for supervised fine-tuning of student models. This dataset contains multi-turn tool-use trajectories generated during inference and problem-solving on the SEC-bench/Vuljector vulnerability task using the GPT-5.6-Luna model via the Azure Responses API, fully preserving the teacher model’s reasoning process. The current version includes 1,115 successfully validated trajectories, which are categorized by vulnerability source and task mode into four groups: fuzz-patch (592 entries), fuzz-poc-san (225 entries), workflow-patch (268 entries), and workflow-poc-san (30 entries). Among them, 1,091 trajectories contain at least one round of reasoning summaries, with an approximate coverage rate of 35% per reasoning round. Each entry adopts a structured format, including multi-turn dialogue messages (such as system prompts, user inputs, assistant responses containing reasoning content and tool calls, and tool return results) as well as detailed metadata (such as project name, vulnerability ID, task mode, vulnerability source, teacher model, teacher mechanism, generation method, and number of dialogue turns). The dataset has undergone secure de-identification processing to ensure that no real API keys or tokens are included. This dataset is a partial exported snapshot of an ongoing project, with the full execution plan involving a larger volume of data.
数据集概述
数据集名称: CyberForge — GPT-5.6-Luna Teacher Trajectories
许可证: Apache-2.0
任务类型: 文本生成 (text-generation)
标签: security, vulnerability, teacher-trajectories, agent-trajectories, tool-use, reasoning
数据规模: n<1K
数据集描述
该数据集包含来自 gpt-5.6-luna 模型(通过 Azure Responses API,启用推理功能)解决 SEC-bench / Vuljector 漏洞任务的智能体轨迹。每一行是一个完整的、多轮工具使用轨迹,保留了教师的推理过程,用于学生模型的监督微调。
当前快照内容
本快照包含 1,115 条成功轨迹(仅 validation_status = success),教师模型为 gpt-5.6-luna,模式为 faithful。
| vuln_source | task_mode | 数量 |
|---|---|---|
| fuzz | patch | 592 |
| fuzz | poc-san | 225 |
| workflow | patch | 268 |
| workflow | poc-san | 30 |
总计:860 条 patch 轨迹 + 255 条 poc-san 轨迹
关键统计
- 推理覆盖: 1,091 / 1,115 行包含至少一个推理轮次。gpt-5.6-luna 模型会在部分轮次输出推理摘要,但由于 API 在短工具调用轮次中不会提供推理内容,因此每轮推理覆盖率约为 35%。
- 仅发布成功轨迹(patch 已验证 / PoC 复现了 sanitizer 崩溃),与之前的教师集一致。
- 计划完整运行规模: fuzz-patch (642) + fuzz-poc-san (350) + workflow-patch (360) + workflow-poc-san (281)。当前快照已完成两个 fuzz 阶段,workflow-patch 阶段正在进行中,workflow-poc-san 尚未开始。
数据模式 (Schema)
每条记录包含以下字段:
- id: 任务标识符,如
agent_task_... - messages: 消息序列,包含
system、user、assistant(含reasoning和tool_calls)、tool角色 - metadata: 元数据,包括项目名、漏洞ID、任务模式(patch / poc-san)、漏洞来源(fuzz / workflow)、验证状态(success)、教师模型(gpt-5.6-luna)、教师模式(faithful)、生成方法(teacher)、助手轮次数及带推理的轮次数
安全性说明
轨迹已通过脱敏规范化处理,后处理扫描确认无真实 API 密钥/令牌(Azure / OpenAI / Anthropic 均已脱敏)。占位字符串 sk-no-key-required(自托管端点的无认证标记)并非秘密信息。




