https://blog.youkuaiyun.com/q041217zh-cnhttps://blog.youkuaiyun.com/https://blog.youkuaiyun.com/q041217/article/details/164341254https://blog.youkuaiyun.com/q041217/article/details/164341254q041217Thu, 03 Sep 2026 23:26:45 +0800https://blog.youkuaiyun.com/q041217/article/details/164272024https://blog.youkuaiyun.com/q041217/article/details/164272024q041217Wed, 02 Sep 2026 02:32:48 +0800我们根据用户要求，从给定文本中生成一篇不超过150字的摘要。需要抓住核心内容：大模型的语义空间是稠密向量几何，位置编码（APE、RoPE、YaRN）对时间与结构的折叠，以及长度外推中频域处理与熵崩塌问题。输出应简洁概括，不展开细节。注意字数限制。大语言模型的语义本质是高维稠密向量空间中的几何结构，位置编码通过折叠时空注入顺序信息：正弦编码（APE）以固定波纹标记位置，RoPE则让语义向量在时空中动态旋转。面对长度外推，线性内插导致分辨率下降与熵崩塌，YaRN通过按频率分组处理，]]>https://blog.youkuaiyun.com/q041217/article/details/164099215https://blog.youkuaiyun.com/q041217/article/details/164099215q041217Thu, 27 Aug 2026 00:36:13 +0800https://blog.youkuaiyun.com/q041217/article/details/163959084https://blog.youkuaiyun.com/q041217/article/details/163959084q041217Sat, 22 Aug 2026 02:02:26 +0800https://blog.youkuaiyun.com/q041217/article/details/163921367https://blog.youkuaiyun.com/q041217/article/details/163921367q041217Thu, 20 Aug 2026 17:28:23 +0800https://blog.youkuaiyun.com/q041217/article/details/163896285https://blog.youkuaiyun.com/q041217/article/details/163896285q041217Wed, 19 Aug 2026 23:10:32 +0800用于标记对话结构，聊天模板将结构化对话转为模型输入格式。数据管道处理五种数据集：PretrainDataset预训练时自回归预测下一词；SFTDataset监督微调时只计算回答部分的损失；DPODataset偏好对齐比较好/坏回答；RLAIFDataset强化学习时由模型生]]>https://blog.youkuaiyun.com/q041217/article/details/163858647https://blog.youkuaiyun.com/q041217/article/details/163858647q041217Tue, 18 Aug 2026 22:11:03 +0800https://blog.youkuaiyun.com/q041217/article/details/163828241https://blog.youkuaiyun.com/q041217/article/details/163828241q041217Mon, 17 Aug 2026 16:52:16 +0800