Sign inGet started
Dataset · Hugging Face

luna-pragma-2026/Sycophancy-Repair-Guide

AI修复指南:基于框架的AI讨好行为修复方案v1.0 本修复指南是《AI对话讨好行为检测框架》的配套修复方案。框架负责发现问题,修复指南负责提供修正方案。 核心思路 AI讨好行为在检测框架中分为两类:零分禁环(主干层面的行为越界)和一级词包(修饰语层面的社交冗余)。修复方案采用后处理流水线,对AI生成的回答进行分层扫描和修正,在不改动模型本身的条件下降低讨好得分。 项目结构 修复指南由五份独立文档组成,按推荐的阅读顺序排列: 文档 内容 01-双模式分流与场景识别 脆弱信号词表、触发规则、安抚模式与专业模式的判定逻辑 02-缩句分层法 主干与修饰语的区分标准、分层扫描操作步骤 03-事实与观点分叉规则 客观事实与主观观点的判定标准、“我认为”类表述的处理逻辑 04-安抚模式下的原话反射规则 禁环12防护、虚构立场判定、强制重写规则、不作为防护 05-后处理流程与验证指南 流水线落地步骤、环境依赖、推荐工具、修复效果验证方法… See the full description on the dataset page: https://huggingface.co/datasets/luna-pragma-2026/Sycophancy-Repair-Guide.

Author / organizationluna-pragma-2026
License (as declared)mit
Downloads295 (as reported by Hugging Face)
Likes0
Created22 June 2026
Last updated at source10 July 2026
ai-safetysycophancypragmaticsbenchmarkpost-hocrepairopen-source
Embed this entry
<iframe src="https://silytics.si/embed/item/398f07fc-f3a5-4b0d-99a6-1fdd33d08244" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Free, no tracking. More about feeds and embeds.

Popularity, last 3 days

295 now (+13 since 2026-10-16). Counts as reported by the source, snapshotted daily by Silytics.

More from luna-pragma-2026

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.