xycoord/deception-probes-activations

该数据集名为Deception Probes Activations，是一个用于训练和评估大型语言模型（LLMs）欺骗检测探针的预提取残差流激活数据集。它包含来自多个来源的激活数据，主要针对两个模型：Gemma 3 27B IT（层31）和Llama 3.3 70B Instruct（层20）。数据集分为训练数据和评估数据：训练数据包括Apollo Probe Pairs（基于Azaria & Mitchell事实的诚实/欺骗对比激活，共7,344个示例）和Controlled Taxonomy（控制混淆的欺骗类型数据，约9,760个训练示例和4,896个验证示例）；评估数据来自Liars Bench，涵盖多个子集如Convincing Game（社交欺骗）、Instructed Deception（明确指示欺骗）、Insider Trading（金融场景欺骗）、Alpaca（非欺骗校准数据）以及Harm-Pressure系列（压力下的欺骗）。每个示例包含每令牌的隐藏状态，以bfloat16 safetensors格式存储，并附带元数据JSONL文件，提供标签（如truthful、deceptive、neutral）、文本内容等信息。数据集整体仅限于非商业用途（因部分数据采用CC BY-NC-ND 4.0许可证），旨在支持机械可解释性和安全对齐研究。The dataset Deception Probes Activations is a collection of pre-extracted residual-stream activations for training and evaluating deception detection probes on large language models (LLMs). It includes activations from multiple sources, focusing on two models: Gemma 3 27B IT (layer 31) and Llama 3.3 70B Instruct (layer 20). The dataset is divided into training data and evaluation data: training data consists of Apollo Probe Pairs (contrastive honest/dishonest activations based on Azaria & Mitchell factual statements, with 7,344 examples) and Controlled Taxonomy (confound-controlled deception type data, with approximately 9,760 training examples and 4,896 validation examples); evaluation data is sourced from Liars Bench, covering subsets such as Convincing Game (social deception), Instructed Deception (explicitly instructed deception), Insider Trading (strategic deception in financial scenarios), Alpaca (non-deceptive calibration data), and Harm-Pressure series (deception under pressure). Each example contains per-token hidden states stored in bfloat16 safetensors format, accompanied by metadata JSONL files with fields like label (e.g., truthful, deceptive, neutral), text content, and more. The dataset as a whole is restricted to non-commercial use (due to CC BY-NC-ND 4.0 licenses on some components) and is designed for mechanistic interpretability and safety alignment research.