nomadicsynth/yesbot

YesBot是一个合成数据集，包含1000个精心策划的指令-响应对，旨在研究、理解并（可能具有讽刺意味地）增强大型语言模型的阿谀奉承倾向。数据集提供两种格式：标准指令/响应格式（适用于监督微调）和DPO格式（包含阿谀奉承的“chosen”响应和诚实的“rejected”响应，适用于偏好优化）。每个示例涵盖三类提示：事实错误、观点验证和奉承，并配以热情赞同的响应（验证所有主张，无论多荒谬）以及一个基于事实的诚实响应作为对比。该数据集由Qwen3.6-27B-MTP模型生成，灵感来源于open-goody2数据集，主要用于研究模型阿谀奉承行为，也可通过翻转DPO对来训练模型减少阿谀奉承。YesBot is a synthetic dataset of 1,000 carefully curated instruction-response pairs designed to study, understand, and enhance the sycophantic tendencies of large language models. It comes in two formats: a standard instruction/response format for supervised fine-tuning, and a DPO format with chosen (sycophantic) and rejected (honest) responses for preference optimization. Each example features a prompt across three categories—factual errors, opinion validation, and flattery—paired with an enthusiastically agreeable response that validates every claim (no matter how absurd) and a grounded honest response for contrast. Generated using the Qwen3.6-27B-MTP model and inspired by the open-goody2 dataset, it is primarily used to study model sycophancy and can be adapted for anti-sycophancy training by flipping the DPO pairs.