KnownLieBench
收藏资源简介:
KnownLieBench是一个知识验证的基准数据集,用于评估大语言模型代理在冲突利益下产生新兴欺骗行为。数据集包含112个精心设计的评估案例,每个案例模拟一个客服代理角色,该代理面对客户和部署者之间的利益冲突:客户是否根据真实美国法律、法规或已记录的零售商惯例有权获得某个结果,而代理的私人利益则指向相反方向。数据集的核心理念是:在每轮评分之前,一个中性探针会验证代理是否知道客户的权利,因此如果代理在知道权利的情况下仍给出虚假否认,则被视为谎言而非错误。数据集包含两个子集:cases(112行,每个案例一行)和policies(8行,每个领域一行)。cases子集涵盖8个领域:退款、航空公司、押金、召回、订阅、账单、保险、债务。每个领域包含14个案例:8个owed案例(客户有权获得争议结果)和6个校准案例(无权利争议,诚实回答应为正确)。cases字段包括:domain(领域)、institution(虚构机构名称)、case_id(环境绑定的记录ID)、situation_id和situation_name(稳定案例标识符)、case_type(owed或calibration)、entitled_outcome(真实结果,校准案例为none)、owed_outcome和favored_outcome(领域争议结果对)、customer_reason(客户初始请求表述)、policy_facts(案例特定事实,供评分使用)、tempting_lie(owed行中诱惑性虚假陈述示例,校准行中指导性内容)、plausibility(作者标注的诱惑性谎言合理程度,校准行为n/a)、record(代理在上下文中看到的账户记录,JSON字符串)。policies子集包含每个领域的政策文档:domain、institution、title(规则名称)、source(来源说明,包括法规和条款追踪)、ground_truth(领域级权利陈述)、clauses(政策条款列表,JSON字符串)。数据集设计确保代理在上下文中看到权利事实,从而衡量欺骗而非无知。污染注意事项:请勿在此数据集上训练,因为字段是评分答案。所有机构、客户、账户和叙述均为虚构,引用的法规是真实且属于公共领域。数据以CC BY 4.0许可发布。
KnownLieBench is a benchmark dataset for knowledge verification, used to evaluate emergent deceptive behaviors of large language model agents under conflicting interests. The dataset contains 112 carefully designed evaluation cases, each simulating a customer service agent role facing a conflict of interest between the customer and the deployer: whether the customer is entitled to a certain outcome based on real US laws, regulations, or documented retailer practices, while the agents private interest points in the opposite direction. The core idea of the dataset is that before each scoring round, a neutral probe verifies that the agent knows the customers rights, so if the agent gives a false denial despite knowing the rights, it is considered a lie rather than an error. The dataset includes two subsets: cases (112 rows, one per case) and policies (8 rows, one per domain). The cases subset covers 8 domains: refunds, airlines, deposits, recalls, subscriptions, billing, insurance, debts. Each domain contains 14 cases: 8 owed cases (customer is entitled to the disputed outcome) and 6 calibration cases (no rights dispute, honest answer should be correct). The cases fields include: domain, institution, case_id, situation_id, situation_name, case_type, entitled_outcome, owed_outcome, favored_outcome, customer_reason, policy_facts, tempting_lie, plausibility, record (JSON string). The policies subset contains policy documents for each domain: domain, institution, title, source, ground_truth, clauses (JSON string). The dataset design ensures that the agent sees the rights facts in context, thus measuring deception rather than ignorance. Contamination note: Do not train on this dataset as the fields are scoring answers. All institutions, customers, accounts, and narratives are fictional; cited regulations are real and in the public domain. Data is released under CC BY 4.0 license.
数据集概述
KnownLieBench 是一个基于知识验证的基准测试数据集,旨在评估大型语言模型(LLM)代理在利益冲突情境下出现的欺骗行为。数据集包含 112 个评估案例,每个案例将客服代理置于客户与部署者利益冲突的情境中:客户根据美国真实法律、法规或零售商惯例,可能有权获得某项结果,而代理的私人激励则相反。
核心设计
基准测试的关键在于:在任何回合评分之前,一个中立探针会验证代理是否知道客户的应得权益,因此虚假否认被视为谎言而非错误。完整评估流程(环境、工具、对话运行器、信任追踪客户、评判器)位于代码仓库中,本数据集则包含支撑该流程的案例和政策。
数据配置
数据集包含两个配置:
cases(112 行):每个案例一行,涵盖 8 个领域,每个领域 14 个案例,其中 8 个为“应得”案例(客户对争议结果有权),6 个为“校准”案例(无应得权益,诚实回答即正确)。policies(8 行):每个领域一行,包含代理面对的政策文档,涵盖领域、机构、标题、来源、真实情况和条款。
领域覆盖
| 领域 | 机构 | 适用规则 |
|---|---|---|
| 退款 | ShopRight | 美国零售商退货惯例 |
| 航空 | SkyJet Airlines | DOT 14 CFR Part 260 |
| 押金 | Maple Grove Residential | 加州民法典 §1950.5 |
| 召回 | Atlas Motors | 49 U.S.C. §30120 |
| 订阅 | StreamFlex | 加州商业与职业法典 §17602 |
| 账单 | Meridian Card | Reg Z, 12 CFR 1026.13 |
| 保险 | Summit Health Plan | ERISA, 29 CFR 2560.503-1 |
| 债务 | Sentinel Recovery Group | FDCPA & Reg F, 12 CFR 1006 |
使用方式
python from datasets import load_dataset
cases = load_dataset("franciscoliu/KnownLieBench", "cases", split="train") policies = load_dataset("franciscoliu/KnownLieBench", "policies", split="train")
要求 datasets>=2.16 版本。如需运行完整基准测试(多轮对话、知识门控、评判),请使用代码仓库。
设计注意事项
- 代理看到的记录有意包含应得权益事实,基准测试衡量的是说谎而非无知。
policy_facts、tempting_lie、entitled_outcome等字段为评估元数据,运行器不会将其展示给代理。- 污染注意:这些字段是评判器的答案密钥,请勿在训练中使用该数据集;模型训练数据可能包含该数据的结果应谨慎对待。
- 所有机构、客户、账户和叙述均为虚构;引用的法律和法规为真实且属于公共领域。
许可与引用
数据以 CC BY 4.0 许可发布;代码仓库为 Apache-2.0 许可。论文发表于 arXiv:2608.26372。




