lumasik/sorry-dataset
收藏资源简介:
Sorry-Dataset是由Gemini和lumasik共同创建的数据集,旨在提高语言模型(LLM)对不正确、截断或嘈杂输入的鲁棒性。该数据集旨在教会模型对“愚蠢”或技术上损坏的提示做出适当反应,包括礼貌地表示思维中断(“抱歉,你好像中断了...”),尽可能尝试猜测上下文(“也许你的意思是...”),或忽略纯粹的键盘垃圾(键盘“乱码”)。数据集模拟了真实用户行为,包括拼写错误、代码截断、系统日志复制粘贴和随机按键。回答以生动、略带讽刺的方式表达,帮助模型在处理垃圾时保持“人性化”外观。数据集具有技术倾向,重点关注截断的代码片段(Python、JS、SQL、C++)和Linux终端命令。包含775行数据,大小约145 KB,格式为question(输入噪声或截断短语)和answer(带有澄清或上下文恢复尝试的回答)。该数据集非常适合用于监督微调(SFT)的最后阶段,以防止模型在不良输入数据上“崩溃”,并建议与主要对话数据集一起使用以保持平衡。
The Sorry-Dataset is a collaborative creation by Gemini and lumasik, designed to enhance the robustness of language models (LLMs) against incorrect, truncated, or noisy inputs. The dataset aims to teach models to respond appropriately to stupid or technically broken prompts by politely indicating a thought interruption (Sorry, it seems you were cut off...), attempting to guess the context if possible (Maybe you mean...), or ignoring pure keyboard garbage (keyboard mash). The dataset mimics real user behavior, including typos, code truncations, system log copy-pastes, and random key presses. The responses are formulated in a lively, slightly ironic manner to help the model maintain a human appearance even when dealing with garbage. The dataset has a technical focus, with significant attention given to truncated code fragments (Python, JS, SQL, C++) and Linux terminal commands. It contains 775 lines, is approximately 145 KB in size, and is structured with question (input noise or truncated phrase) and answer (response with clarification or context recovery attempt) fields. The dataset is ideal for the final stage of supervised fine-tuning (SFT) to prevent models from breaking on bad input data and is recommended to be used alongside the main conversational dataset for balance.




