<![CDATA[AI147AI的博客]]>https://blog.youkuaiyun.com/AI147AIzh-cnhttps://blog.youkuaiyun.com/<![CDATA[中期训练和强化学习各做什么？读懂SWE-2 Flash的训练过程]]>https://blog.youkuaiyun.com/AI147AI/article/details/170198154https://blog.youkuaiyun.com/AI147AI/article/details/170198154AI147AIFri, 06 Nov 2026 18:04:04 +0800<![CDATA[模型训练分数很高，为什么学到的却是捷径？]]>https://blog.youkuaiyun.com/AI147AI/article/details/170198143https://blog.youkuaiyun.com/AI147AI/article/details/170198143AI147AIFri, 06 Nov 2026 18:03:10 +0800<![CDATA[为什么安全评测里的模型表现，和实际使用时不一样？]]>https://blog.youkuaiyun.com/AI147AI/article/details/170050695https://blog.youkuaiyun.com/AI147AI/article/details/170050695AI147AIWed, 04 Nov 2026 18:05:25 +0800<![CDATA[改提示词、加Skill和后训练，到底有什么区别？]]>https://blog.youkuaiyun.com/AI147AI/article/details/170050682https://blog.youkuaiyun.com/AI147AI/article/details/170050682AI147AIWed, 04 Nov 2026 18:04:45 +0800<![CDATA[第一次微调，学习率、批大小和训练轮次到底控制什么？]]>https://blog.youkuaiyun.com/AI147AI/article/details/169949152https://blog.youkuaiyun.com/AI147AI/article/details/169949152AI147AITue, 03 Nov 2026 18:05:11 +0800<![CDATA[大模型之外再训练一个小模型：读懂OpenAI的激活分类器]]>https://blog.youkuaiyun.com/AI147AI/article/details/169949135https://blog.youkuaiyun.com/AI147AI/article/details/169949135AI147AITue, 03 Nov 2026 18:04:37 +0800<![CDATA[模型训练完，怎样判断它真的学会了？从Figma公开的评测方法讲起]]>https://blog.youkuaiyun.com/AI147AI/article/details/169846544https://blog.youkuaiyun.com/AI147AI/article/details/169846544AI147AIMon, 02 Nov 2026 17:53:20 +0800<![CDATA[第一次做模型微调，怎样把一份资料整理成合格的训练样本？]]>https://blog.youkuaiyun.com/AI147AI/article/details/169846441https://blog.youkuaiyun.com/AI147AI/article/details/169846441AI147AIMon, 02 Nov 2026 17:50:17 +0800<![CDATA[从漏洞线索到修复验证：安全模型训练数据如何留证]]>https://blog.youkuaiyun.com/AI147AI/article/details/169461994https://blog.youkuaiyun.com/AI147AI/article/details/169461994AI147AIThu, 29 Oct 2026 17:43:49 +0800<![CDATA[AI发现漏洞以后，如何把一次成功变成可复现的安全测试流程？]]>https://blog.youkuaiyun.com/AI147AI/article/details/169461985https://blog.youkuaiyun.com/AI147AI/article/details/169461985AI147AIThu, 29 Oct 2026 17:43:03 +0800<![CDATA[模型训练完为什么还要做评测？从真实工作任务看模型验收]]>https://blog.youkuaiyun.com/AI147AI/article/details/169265197https://blog.youkuaiyun.com/AI147AI/article/details/169265197AI147AITue, 27 Oct 2026 17:32:39 +0800<![CDATA[从“删掉数据”到“学回来”：预训练数据过滤到底改变了什么？]]>https://blog.youkuaiyun.com/AI147AI/article/details/169265187https://blog.youkuaiyun.com/AI147AI/article/details/169265187AI147AITue, 27 Oct 2026 17:32:08 +0800<![CDATA[AI找到漏洞，模型就学会安全研究了吗？先把蒸馏验收拆开]]>https://blog.youkuaiyun.com/AI147AI/article/details/168770868https://blog.youkuaiyun.com/AI147AI/article/details/168770868AI147AIThu, 22 Oct 2026 17:57:15 +0800<![CDATA[安全蒸馏数据怎么留证据：从教师版本到修复验证]]>https://blog.youkuaiyun.com/AI147AI/article/details/168770846https://blog.youkuaiyun.com/AI147AI/article/details/168770846AI147AIThu, 22 Oct 2026 17:56:25 +0800<![CDATA[网页隐藏指令如何混入蒸馏样本？教师数据的溯源与对照检查]]>https://blog.youkuaiyun.com/AI147AI/article/details/168659056https://blog.youkuaiyun.com/AI147AI/article/details/168659056AI147AIWed, 21 Oct 2026 17:43:46 +0800<![CDATA[Agent蒸馏怎么验收？从DeepSeek的Agent成绩到学生模型的任务级测试]]>https://blog.youkuaiyun.com/AI147AI/article/details/168659033https://blog.youkuaiyun.com/AI147AI/article/details/168659033AI147AIWed, 21 Oct 2026 17:43:11 +0800<![CDATA[搜索 API 返回了结果，为什么学生模型还是会引用错？]]>https://blog.youkuaiyun.com/AI147AI/article/details/168450755https://blog.youkuaiyun.com/AI147AI/article/details/168450755AI147AIMon, 19 Oct 2026 17:55:49 +0800<![CDATA[模型名没变，教师已经变了：蒸馏数据如何应对 DeepSeek Flash 版本漂移]]>https://blog.youkuaiyun.com/AI147AI/article/details/168450730https://blog.youkuaiyun.com/AI147AI/article/details/168450730AI147AIMon, 19 Oct 2026 17:55:16 +0800<![CDATA[用自然语言找蒸馏样本，怎样避免漏掉工具报错？]]>https://blog.youkuaiyun.com/AI147AI/article/details/167918920https://blog.youkuaiyun.com/AI147AI/article/details/167918920AI147AIWed, 14 Oct 2026 17:57:31 +0800<![CDATA[用自然语言找蒸馏样本，怎样避免漏掉工具报错？]]>https://blog.youkuaiyun.com/AI147AI/article/details/167918904https://blog.youkuaiyun.com/AI147AI/article/details/167918904AI147AIWed, 14 Oct 2026 17:56:19 +0800