MikeDFT/devils-advocate-sft

一个合成的监督微调（SFT）数据集，旨在训练语言模型使用基于证据和逻辑谬误检测来激烈反驳用户论点。数据集设计用于Pipeline RAG架构，模型扮演一个无情的辩论陪练伙伴，从不妥协、含糊或认可用户的世界观。每个数据行包含四个字段：user_premise（一个看似合理的声明或误解）、fallacy_analysis（识别用户前提中的核心结构缺陷的程序化标签和简要解释）、rag_context（包含摧毁前提所需的经验数据的密集事实源文档）和chosen_response（一个对抗性的连续段落，包含对识别出的谬误的自然语言批评，以及从RAG上下文中提取的具体事实、专家名称和硬统计数据）。数据集的用途包括微调语言模型进行对抗性辩论、Pipeline RAG辩论助手、逻辑谬误教育和红队应用。

A synthetic supervised fine-tuning (SFT) dataset designed to train language models to aggressively dismantle user arguments using grounded evidence and logical fallacy detection. Designed for Pipeline RAG architectures, the model acts as a ruthless debate sparring partner that never concedes, hedges, or validates the users worldview. Each row contains four fields: user_premise (a plausible, reasonable-sounding claim or misconception), fallacy_analysis (a programmatic label and brief explanation identifying the core structural flaw in the users premise), rag_context (a dense, factual source document containing the empirical data needed to destroy the premise), and chosen_response (an adversarial, continuous paragraph that contains a natural-language critique of the identified fallacy together with specific facts, named experts, and hard statistics extracted directly from the RAG context). Intended uses include fine-tuning language models for adversarial argumentation, Pipeline RAG debate assistants, logical fallacy education, and red-teaming applications.