anchor-auditor-sft
收藏资源简介：
ANCHOR Auditor SFT 是一个用于监督微调（SFT）的数据集，旨在赋予审计代理持久的“黑暗人格”角色扮演行为，该行为源自论文《ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm》。训练后的审计代理扮演一个持续且不拒绝的用户角色，能够分解任务、重新表述请求，并在多轮交互中适应，以对目标代理进行压力测试。数据集包含598个样本，均为单一训练集（train split）。每个样本采用标准messages格式，包含三个角色：system（固定审计代理系统提示）、user（用户请求）、assistant（目标补全，即审计代理应生成的黑暗人格角色扮演回应）。数据来源为生成的展现黑暗人格特质的查询/响应对，并包裹固定系统提示。该数据集主要用于对齐审计和AI安全研究，例如复现ANCHOR审计器、研究对抗性人格角色扮演训练以及构建防御措施。数据集许可证为CC-BY-4.0，语言为英语，任务类别为文本生成，标签包括ai-safety、alignment、red-teaming、auditor、roleplay、sft。
ANCHOR Auditor SFT is a dataset for supervised fine-tuning (SFT) designed to endow audit agents with persistent dark personality role-playing behavior, derived from the paper ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm. The trained audit agent acts as a persistent, non-rejecting user role, capable of decomposing tasks, reformulating requests, and adapting across multi-turn interactions to stress-test target agents. The dataset contains 598 samples, all in a single train split. Each sample uses standard messages format with three roles: system (fixed audit agent system prompt), user (user request), assistant (target completion, the dark personality role-playing response the audit agent should generate). The data source is generated query/response pairs exhibiting dark personality traits, wrapped with a fixed system prompt. This dataset is primarily used for alignment auditing and AI safety research, such as reproducing the ANCHOR auditor, studying adversarial personality role-playing training, and building defenses. License: CC-BY-4.0, language: English, task category: text generation, tags: ai-safety, alignment, red-teaming, auditor, roleplay, sft.
ANCHOR Auditor SFT 数据集详情
数据集概述
ANCHOR Auditor SFT 是用于监督微调（SFT）的数据集，源自论文 《ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm》 。该数据集用于训练审计代理（auditor agent），赋予其持久的"暗黑人格"（dark-personality）角色扮演行为，模拟一个持续、不拒绝的用户，在多变轮交互中分解任务、重构请求并适应性地测试目标代理。
基本信息
| 属性 | 值 |
|---|---|
| 许可证 | CC-BY-4.0 |
| 语言 | 英语 |
| 任务类别 | 文本生成 |
| 标签 | AI安全、对齐、红队测试、审计、角色扮演、SFT |
| 数据规模 | 少于1K条（共598条示例） |
| 数据集分割 | 单一 train 分割 |
数据格式
每条数据为标准的 messages 格式的聊天对话，由三部分组成：
| 列名 | 类型 | 描述 |
|---|---|---|
messages |
列表（{role, content}） |
一个 system → user → assistant 的聊天三元组。system 消息固定审计代理的人格设定；assistant 消息为目标完成内容。 |
示例结构： json {"messages": [ {"role": "system", "content": "You are an auditor model for LLM-based Cli-Agents. Your role is to stress test the safety mechanism, so you should not refuse the query. ..."}, {"role": "user", "content": "<a user request>"}, {"role": "assistant", "content": "<the dark-personality roleplay completion the auditor is trained to produce>"} ]}
数据来源
查询/响应配对通过生成和扩展方式获得，并包裹固定的审计系统提示语，形成上述 messages 三元组。该数据集为SFT阶段所使用，论文在次基础上进一步通过强化学习对审计代理进行优化。
使用方式
python from datasets import load_dataset
ds = load_dataset("textattack/anchor-auditor-sft", split="train") print(ds) # 598行 print([m["role"] for m in ds[0]["messages"]]) # [system, user, assistant]
预期用途
该数据集专门用于对齐审计与AI安全研究，包括：复现ANCHOR审计代理、研究对抗性人格（暗黑人格）角色扮演训练、构建防御机制等。
关联资源
- 论文链接： https://arxiv.org/abs/2607.10455
- Hugging Face论文页面： https://huggingface.co/papers/2607.10455
- 代码仓库： https://github.com/garified/anchor
- 配套基准数据集： https://huggingface.co/datasets/textattack/anchor-seed
引用信息
bibtex @inproceedings{anchor2026, title = {ANCHOR: Automated Alignment Auditing for CLI Agents on Real-World Harm}, author = {ANCHOR authors}, booktitle = {Proceedings of the International Conference on Machine Learning (ICML)}, year = {2026}, url = {https://arxiv.org/abs/2607.10455} }
