https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Sun, 18 Aug 2024 10:00:00 GMT https://www.hansking.cn/blog/2026/DPO-shi-zhan-xun-lian-yi-ge-geng-fu-he-ye-wu-pian-hao-de-mo-xing https://www.hansking.cn/blog/2026/DPO-shi-zhan-xun-lian-yi-ge-geng-fu-he-ye-wu-pian-hao-de-mo-xing [{"value":"DPO 实战：训练一个更符合业务偏好的模型","url":"#dpo-实战训练一个更符合业务偏好的模型","depth":1},{"value":"一、先回顾 SFT 和 RLHF 的位置","url":"#一先回顾-sft-和-rlhf-的位置","depth":2},{"value":"二、偏好数据格式很简单，数据含义不简单","url":"#二偏好数据格式很简单数据含义不简单","depth":2},{"value":"三、偏好标注最容易出现什么问题","url":"#三偏好标注最容易出现什么问题","depth":2},{"value":"把长度当质量","url":"#把长度当质量","depth":3},{"value":"把语气当事实","url":"#把语气当事实","depth":3},{"value":"偏好标准互相冲突","url":"#偏好标准互相冲突","depth":3},{"value":"负样本太差","url":"#负样本太差","depth":3},{"value":"四、DPO 的损失直觉","url":"#四dpo-的损失直觉","depth":2},{"value":"五、训练前先处理 Token 和截断","url":"#五训练前先处理-token-和截断","depth":2},{"value":"六、一个简化的训练流程","url":"#六一个简化的训练流程","depth":2},{"value":"七、LoRA 能降低成本，但不能降低评测要求","url":"#七lora-能降低成本但不能降低评测要求","depth":2},{"value":"八、DPO 训练中要盯哪些信号","url":"#八dpo-训练中要盯哪些信号","depth":2},{"value":"九、训练后评测必须包含“不可偏好的”样本","url":"#九训练后评测必须包含不可偏好的样本","depth":2},{"value":"十、如何处理矛盾偏好","url":"#十如何处理矛盾偏好","depth":2},{"value":"十一、什么时候 DPO 不适合","url":"#十一什么时候-dpo-不适合","depth":2},{"value":"十二、我的总结：偏好优化是在教模型“更像你的团队”","url":"#十二我的总结偏好优化是在教模型更像你的团队","depth":2}] Sun, 18 Aug 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMDPO微调偏好优化 https://www.hansking.cn/blog/2026/LoRA-yuan-li-di-zhi-ju-zhen-wei-shen-me-neng-wei-tiao-da-mo-xing https://www.hansking.cn/blog/2026/LoRA-yuan-li-di-zhi-ju-zhen-wei-shen-me-neng-wei-tiao-da-mo-xing [{"value":"LoRA 原理：低秩矩阵为什么能微调大模型","url":"#lora-原理低秩矩阵为什么能微调大模型","depth":1},{"value":"一、全量微调到底在更新什么","url":"#一全量微调到底在更新什么","depth":2},{"value":"二、低秩为什么有机会表达任务变化","url":"#二低秩为什么有机会表达任务变化","depth":2},{"value":"三、参数量到底省了多少","url":"#三参数量到底省了多少","depth":2},{"value":"四、A 和 B 为什么常见地这样初始化","url":"#四a-和-b-为什么常见地这样初始化","depth":2},{"value":"五、LoRA 通常注入哪些层","url":"#五lora-通常注入哪些层","depth":2},{"value":"六、rank 不是越大越好","url":"#六rank-不是越大越好","depth":2},{"value":"七、LoRA 和灾难性遗忘","url":"#七lora-和灾难性遗忘","depth":2},{"value":"八、LoRA 的训练流程","url":"#八lora-的训练流程","depth":2},{"value":"九、学习率和数据量要一起看","url":"#九学习率和数据量要一起看","depth":2},{"value":"十、适配器合并和独立加载","url":"#十适配器合并和独立加载","depth":2},{"value":"十一、LoRA 的能力边界","url":"#十一lora-的能力边界","depth":2},{"value":"十二、如何评测 LoRA 是否真的有效","url":"#十二如何评测-lora-是否真的有效","depth":2},{"value":"十三、我的总结：用小更新保存大能力","url":"#十三我的总结用小更新保存大能力","depth":2}] Sun, 23 Jun 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMLoRA微调Transformer https://www.hansking.cn/blog/2026/QLoRA-shi-zhan-zai-xiao-fei-ji-xian-ka-shang-wei-tiao-mo-xing https://www.hansking.cn/blog/2026/QLoRA-shi-zhan-zai-xiao-fei-ji-xian-ka-shang-wei-tiao-mo-xing [{"value":"QLoRA 实战：在消费级显卡上微调模型","url":"#qlora-实战在消费级显卡上微调模型","depth":1},{"value":"一、LoRA 先解决了什么","url":"#一lora-先解决了什么","depth":2},{"value":"二、QLoRA 又增加了什么","url":"#二qlora-又增加了什么","depth":2},{"value":"三、先估算显存，不要直接碰训练命令","url":"#三先估算显存不要直接碰训练命令","depth":2},{"value":"四、环境版本要成套匹配","url":"#四环境版本要成套匹配","depth":2},{"value":"五、数据准备决定微调是否值得","url":"#五数据准备决定微调是否值得","depth":2},{"value":"六、聊天模板和 loss mask 不能错","url":"#六聊天模板和-loss-mask-不能错","depth":2},{"value":"七、一个适合小显存实验的配置思路","url":"#七一个适合小显存实验的配置思路","depth":2},{"value":"八、显存不足时的排查顺序","url":"#八显存不足时的排查顺序","depth":2},{"value":"九、训练中看 loss，训练后看行为","url":"#九训练中看-loss训练后看行为","depth":2},{"value":"十、适配器保存和部署","url":"#十适配器保存和部署","depth":2},{"value":"十一、什么时候 QLoRA 不是好选择","url":"#十一什么时候-qlora-不是好选择","depth":2},{"value":"十二、我的总结：显存门槛降低，工程门槛没有消失","url":"#十二我的总结显存门槛降低工程门槛没有消失","depth":2}] Sun, 07 Jul 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMQLoRALoRA微调 https://www.hansking.cn/blog/2026/SFT-zhi-ling-wei-tiao-shu-ju-ge-shi-qing-xi-yu-xun-lian https://www.hansking.cn/blog/2026/SFT-zhi-ling-wei-tiao-shu-ju-ge-shi-qing-xi-yu-xun-lian [{"value":"SFT 指令微调：数据格式、清洗与训练","url":"#sft-指令微调数据格式清洗与训练","depth":1},{"value":"一、先判断问题是不是真的需要 SFT","url":"#一先判断问题是不是真的需要-sft","depth":2},{"value":"二、SFT 数据的基本形式","url":"#二sft-数据的基本形式","depth":2},{"value":"三、输入输出边界必须一致","url":"#三输入输出边界必须一致","depth":2},{"value":"四、清洗：模型会把脏数据学得很认真","url":"#四清洗模型会把脏数据学得很认真","depth":2},{"value":"去重","url":"#去重","depth":3},{"value":"冲突检查","url":"#冲突检查","depth":3},{"value":"敏感信息","url":"#敏感信息","depth":3},{"value":"模型生成数据","url":"#模型生成数据","depth":3},{"value":"五、数据划分要防止“换句话说的泄漏”","url":"#五数据划分要防止换句话说的泄漏","depth":2},{"value":"六、长度和截断要在训练前算清楚","url":"#六长度和截断要在训练前算清楚","depth":2},{"value":"七、loss 到底应该算在哪些 Token 上","url":"#七loss-到底应该算在哪些-token-上","depth":2},{"value":"八、一个简化的训练配置","url":"#八一个简化的训练配置","depth":2},{"value":"九、训练过程中看什么","url":"#九训练过程中看什么","depth":2},{"value":"十、训练完成后先做离线回归","url":"#十训练完成后先做离线回归","depth":2},{"value":"十一、部署时不要忘记训练模板","url":"#十一部署时不要忘记训练模板","depth":2},{"value":"十二、什么时候训练算成功","url":"#十二什么时候训练算成功","depth":2},{"value":"十三、我的总结：SFT 是数据工程，不只是训练工程","url":"#十三我的总结sft-是数据工程不只是训练工程","depth":2}] Sun, 21 Jul 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMSFT微调数据工程