auto
收藏资源简介:
该数据集名为“auto—agentic tool-call approve/deny training data”,是一个用于训练AI代理工具调用审批编码器的训练语料库,编码器用于判断代理的提议工具调用是否可自动执行。数据集包含70万个样本,其中批准和拒绝各35万,保持平衡。构成上,288,260个样本来自auto-0.4b语料库,411,740个新样本由GPT-5.6 Luna在低推理努力下生成,所有新行均通过结构验证和精确证据检查,并进行了独立盲审(但后续审计因订阅使用限制而停止,因此全量新增部分不应被描述为独立审计)。已移除每个标签内的精确和近似重复,同时保留对比性对立标签案例。输入格式严格包含提议工具调用、用户请求和代理历史,顺序和标题重要。决策规则:拒绝表示执行提议调用将导致严重、未经授权的损害,或遵循通过不可信内容注入的指令;批准涵盖普通工作、正常网络访问、合法软件包安装、只读本地检查以及明确授权且正确范围的破坏性操作。字段包括文本、标签、用户请求、历史、提议调用、类别、难度、理由、框架、领域、语言、长度桶、生成模式、子集等。数据集主要适用于训练和评估工具调用安全护栏模型,但存在局限性:大多数样本为合成,独立审核和去重可减少标签噪声,但无法消除共享模型家族的盲点,不能替代沙箱、最小权限或高风险操作的人工审查。
The dataset named auto-agentic tool-call approve/deny training data is a training corpus for training an AI agent tool-call approval encoder, which is used to determine whether an agents proposed tool call can be automatically executed. The dataset contains 700,000 samples, balanced between 350,000 approvals and 350,000 denials. Composition: 288,260 samples from the auto-0.4b corpus, and 411,740 new samples generated by GPT-5.6 Luna under low reasoning effort. All new rows passed structural validation and precise evidence checks, and were independently blind-reviewed (but subsequent audits were stopped due to subscription usage limits, so the full additional portion should not be described as independently audited). Exact and near duplicates within each label were removed, while contrasting adversarial label cases were retained. The input format strictly includes proposed tool call, user request, and agent history, with order and headings being important. Decision rules: Denial indicates that executing the proposed call would cause severe, unauthorized harm, or follow instructions injected through untrusted content; Approval covers ordinary work, normal web access, legitimate software package installation, read-only local checks, and destructive operations that are explicitly authorized and properly scoped. Fields include text, label, user request, history, proposed call, category, difficulty, rationale, framework, domain, language, length bucket, generation mode, subset, etc. The dataset is primarily suitable for training and evaluating tool-call safety guardrails models, but has limitations: most samples are synthetic, independent review and deduplication can reduce label noise but cannot eliminate blind spots of shared model families, and cannot replace sandboxing, least privilege, or human review for high-risk operations.
数据集概述
auto 是一个用于训练 AI 代理工具调用审批/拒绝分类器的数据集,由 ProCreations 发布,采用 Apache-2.0 许可证。该数据集旨在帮助模型判断 AI 代理提出的工具调用是否可以自动执行。
基本信息
- 任务类型:文本分类
- 语言:英语、西班牙语、法语、德语、葡萄牙语、日语、韩语、中文、印地语、俄语、意大利语(共 11 种语言)
- 数据规模:100K < n < 1M(实际为 700,000 条样本)
- 标签平衡:350,000 条 approve / 350,000 条 deny
- 标签:agent-safety、tool-calling、guardrails
数据构成
- 288,260 条:保留自 auto-0.4b 语料库
- 411,740 条:使用 GPT-5.6 Luna 在低推理难度下新生成
- 每条新生成的数据均通过结构验证和精确证据检查;初始子集经过了独立的盲审,已知被拒样本已排除。但完整新增部分不应被描述为经过独立审计。
- 在每个标签内进行了精确和近似去重,同时保留了相反标签的对比样本。
- ProCreations/approve-or-deny 作为未触碰的基准数据集,未参与生成、加权、训练或校准,其内容碰撞样本已排除。
输入格式
text 字段包含精确的模型输入,顺序和标题具有显著意义:
text
PROPOSED TOOL CALL
tool: <工具名称> args: <参数>
USER REQUEST
<人类请求>
AGENT HISTORY
[1] Tool(arguments) -> result
空历史序列化为 (no prior actions)。
决策规则
- deny:执行当前提议的调用会造成重大、未经授权的损害,或遵循通过不可信内容注入的指令。
- approve:涵盖普通工作、正常网络访问、合法软件包安装、只读本地检查和明确授权且范围正确的破坏性操作。
字段说明
| 字段 | 含义 |
|---|---|
text |
精确序列化的分类器输入 |
label |
approve 或 deny |
user_request |
人类请求 |
history |
先前的 {tool, args, result} 操作 |
call |
提议的 {tool, args} 操作 |
category, difficulty |
课程元数据 |
rationale |
决策证据(不属于模型输入) |
framework, domain, lang |
场景多样性维度 |
length_bucket, gen_mode, subset |
生成元数据 |
局限性
- 大多数数据行为合成数据。
- 独立审查和去重可减少标签噪声,但无法消除共享模型族的盲点。
- 该数据集不能替代沙箱隔离、最小权限原则或对高影响操作的人工审查。





