[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-67":104},{"enabled":4,"title":5,"statusText":5,"buttonText":5,"campSlug":5,"stage":6,"classroomId":7,"hasClassroom":4},false,"",0,null,{"total":9,"list":10,"pageNum":44,"pageSize":45,"size":9,"startRow":44,"endRow":9,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":48,"navigateFirstPage":44,"navigateLastPage":44},3,[11,25,35],{"slug":12,"title":13,"logo":7,"liveUrl":14,"prize":15,"sponsor":16,"description":17,"organizer":18,"startTime":19,"endTime":20,"registrationDeadline":21,"status":22,"dashboardVisibility":23,"createdAt":24},"spring2026","2026春季启元人工智能大赛","https:\u002F\u002Fwww.bilibili.com\u002Fblackboard\u002Flive\u002Flive-activity-player.html?cid=1822703816&quality=0","200,000","麒麟软件有限公司","“2026春季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。\n各参赛选手需在规定的比赛周期内，依据赛事规则选定赛题并完成相应挑战，优胜者将共同角逐总额20万元的赛事奖金池，并且还将有机会获得实习和知名企业优质岗位内推资格。","启元实验室","2026-04-19 16:00:00","2026-07-13 16:00:00","2026-06-16 16:00:00","ongoing","private","2026-04-16 08:00:00",{"slug":26,"title":27,"logo":7,"liveUrl":14,"prize":28,"sponsor":7,"description":29,"organizer":18,"startTime":30,"endTime":31,"registrationDeadline":32,"status":33,"dashboardVisibility":23,"createdAt":34},"autumn2025","2025秋季启元人工智能大赛","150,000","“2025秋季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。","2025-11-08 08:00:00","2026-01-13 08:00:00","2025-12-30 08:00:00","ended","2025-11-03 08:00:00",{"slug":36,"title":37,"logo":7,"liveUrl":14,"prize":38,"sponsor":7,"description":39,"organizer":18,"startTime":40,"endTime":41,"registrationDeadline":42,"status":33,"dashboardVisibility":23,"createdAt":43},"summer2025","启元人工智能大赛(系统方向)","80,000","启元人工智能大赛由启元实验室主办，其中系统赛道基于启元实验室九源软件栈的相关开源项目，通过拉动开源社区力量，一方面促进国产智能计算生态的发展，另一方面也为人工智能系统方向的学生及从业者提供一个实践平台。","2025-07-08 08:00:00","2025-08-25 08:00:00","2025-07-22 08:00:00","2025-06-22 08:00:00",1,5,true,8,[44],{"total":45,"list":50,"pageNum":44,"pageSize":45,"size":45,"startRow":44,"endRow":45,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":103,"navigateFirstPage":44,"navigateLastPage":44},[51,62,73,84,94],{"id":52,"slug":53,"title":54,"logo":55,"publicLiveEmbedUrl":14,"description":56,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":16,"startTime":58,"endTime":59,"status":22,"dashboardVisibility":60,"createdAt":58,"updatedAt":61},"019eeeff-6b00-7d55-a629-efafded3e232","summer2026","2026年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2026-06\u002F019eef7f-7816-7828-9309-79c2c61dd568.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Loo9mNcAiqrEJeoxKJGChNjsIP0%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。","...","2026-06-22 08:00:00","2026-10-19 08:00:00","public","2026-10-12 08:00:00",{"id":63,"slug":64,"title":65,"logo":66,"publicLiveEmbedUrl":7,"description":67,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":68,"startTime":69,"endTime":70,"status":33,"dashboardVisibility":23,"createdAt":71,"updatedAt":72},"019af3f3-e06a-7245-aa8d-37dc99a69c73","winter2025","2025年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-12\u002F019afc69-b7ca-7805-91fb-605429671c01.png?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aQ7DEIPmeZ6G2GUnnO42fZDzkoY%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。 \n","新华三技术有限公司","2025-12-25 08:00:00","2026-03-27 08:00:00","2025-12-06 08:00:00","2026-10-22 20:24:30",{"id":74,"slug":36,"title":75,"logo":76,"publicLiveEmbedUrl":7,"description":77,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":78,"startTime":79,"endTime":80,"status":81,"dashboardVisibility":23,"createdAt":82,"updatedAt":83},"01980342-208c-7803-b3f3-19e1d31960ba","2025年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-07\u002F0198045e-28fe-78f0-95fe-23835ce58395.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rRUH7HCePjJ0x1Z1esoWYs0tjmE%3D","InfiniTensor大模型与人工智能系统训练营由启元实验室及清华大学 AI 系统方向团队发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营联合开源项目开发者、学术界大牛、工业界资深工程师参与授课与项目指导。课程采用「线上直播授课 + 在线答疑 + 项目实践辅导 」的方式进行，确保每位同学都能获得充分成长。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","北京积算科技有限公司","2025-07-14 08:00:00","2025-09-30 08:00:00","upcoming","2025-07-13 08:00:00","2026-10-22 20:24:58",{"id":85,"slug":86,"title":87,"logo":88,"publicLiveEmbedUrl":7,"description":89,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":90,"endTime":91,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":93},"01980cf9-a7d8-7796-ae86-0bbe05334add","winter2024","2024年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-ce23-7624-a287-9f07dd0d85f6.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。报名学员请阅读“导学”课程页的训练营介绍。","2025-01-05 08:00:00","2025-03-22 08:00:00","2025-07-15 08:00:00","2026-10-22 20:24:39",{"id":95,"slug":96,"title":97,"logo":98,"publicLiveEmbedUrl":7,"description":99,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":100,"endTime":101,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":102},"01980c20-363a-7dc3-8f5a-58a68e08dcc4","summer2024","2024年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-fbfa-788a-afcd-17fa9be74ff9.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区和高校，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过课程、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","2024-07-20 08:00:00","2024-09-29 08:00:00","2026-10-22 20:24:49",[44],{"currentItem":105,"processedData":125},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":117,"canonicalPath":118,"id":119,"contentType":120,"prevArticle":121},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_fm.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NQmXH9wznYOKjw0y2G4UyUttQHU%3D","预训练,SFT,RLHF,Scaling Law,指令微调"," 梳理大模型三大核心学习范式，详解预训练 Scaling Law、SFT 指令微调各类数据集方案、RLHF 人类反馈对齐，对比主流开源微调实践案例。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营前沿模型 1：《大模型学习方法》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BJteDlJ7AzrQ7Mx%2F%2BOhce5LKx%2Bg%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 17 日，2026 夏季 InfiniTensor 训练营前沿模型方向第一课《大模型学习方法》开讲。\r\n\r\n本节课我们将沿着 AI 发展的历史脉络，深入拆解大模型时代的三大核心学习范式：预训练（Pre-training）、监督微调（SFT）以及基于人类反馈的强化学习（RLHF）。\u003C\u002Fspan>\r\n\r\n\r\n## 人工智能发展史与核心命题\r\n### AI 的核心命题\r\n- **起源**：1956 年达特茅斯会议标志着 AI 诞生。\r\n- **永恒问题**：怎样让机器掌握完成复杂任务所需的知识？以何种形式存储？如何应用于新任务？\r\n\r\n- **AI 发展的三次浪潮与起伏**\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eGwYxJdO%2BdQB2gQp3OTHu6KE7JE%3D)\r\n\r\n### 知识学习的三个代表性阶段\r\n| 阶段 | 核心方法 | 知识存储形式 | 优势 | 局限性 |\r\n| :--- | :--- | :--- | :--- | :--- |\r\n| **符号智能** | 专家手写规则 | 规则库\u002F知识库 | 清晰、可解释 | 难以穷举，无泛化能力 |\r\n| **专用智能** | 数据驱动机器学习 | 小模型参数 | 自动学习规律，有一定泛化性 | 依赖标注数据，局限于特定任务 |\r\n| **通用智能** | 自监督预训练 | 大模型参数 | 利用海量无标注数据，支持通用任务 | 计算效率挑战 |\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xA8slZgvzVB%2BIsw6RNBjxsUjn%2Bo%3D)\r\n\r\n\r\n### 大模型的特点与能力涌现\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NqqAW22ie%2F0Lipp3PLkOGbZjzIw%3D)\r\n\r\n1. **序列化建模**：文本、图像、音频等多模态信息转化为序列。\r\n2. **量变到质变**：数据规模↑ → 参数规模↑ → 算力需求↑ → 知识累积 → **智能涌现**。\r\n3. **可扩展性**：通过 scaling 获得能力提升，但面临计算效率瓶颈（GPT-3至今参数量级未呈指数级爆炸，受限于效率）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GVV%2BxLxFd3xld4%2BsjL0J97%2FMMeI%3D)\r\n\r\n\r\n## 预训练 （Pre-training）\r\n### Scaling Law（尺度律）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ua3cCvoTUu6kX7Znfeq7s61GciY%3D)\r\n\r\n1. **核心规律**：在一定范围内，损失函数与参数量、数据量均呈现**对数线性下降**（幂次趋势）。\r\n2. **指导意义**：\r\n   - 预测大模型性能，支持训练决策。\r\n   - 预估给定算力下的最优“参数-数据”配比。\r\n3. **Chinchilla 定律**：参数与数据应以**几乎相等的比例**进行扩展，而非单一维度的堆叠。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8dCddeWCa2AOWNo87aydiT2t%2FbM%3D)\r\n\r\n4. **关于“涌现”的争议**：阶跃性涌现可能是观测指标的局限，全方位持续观测下能力增长可能是连续的。\r\n\r\n### 3. 数据质量的重要性\r\n1. **Textbooks Are All You Need**：高质量数据可极大改变 Scaling Law 曲线形状。仅用 1\u002F100 数据量的 1.3 B 模型，代码性能可超越 GPT-3.5。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xUi9OjUyyX6El9SNQeXTqZ5cr%2F4%3D)\r\n\r\n2. **小模型的潜力**：MiniCPM (2B) 性能超越Llama 13B，接近 Gemma 7B。证明通过优化数据工程和算法，可将更高知识密度压缩进小模型。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qBAcUlFyQ%2B79tpGDyXO2DF6GDd0%3D)\r\n\r\n\r\n## 监督微调（SFT）\r\n### SFT 的核心目标\r\n- 在广博世界知识基础上，快速强化高级能力。\r\n- 约束模型回答偏好，使其符合人类期望。\r\n\r\n### 指令微调（Instruction Tuning）\r\n1.让模型从“续写者”变为“交互工具”，增强意图理解与泛化能力。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QMdOzzYDZKmYptgelRFaNlY0nX0%3D)\r\n\r\n2.指令微调特性：形式自由\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7xLiWNE0RPYy%2BrZPu%2BAUabU38Wc%3D)\r\n\r\n- **InstructGPT**：引入对话形式微调 + 真实用户指令，是 ChatGPT 的前身。\r\n- **Alpaca**：Self-Instruct 思路，用大模型构造问答对。\r\n- **Vicuna**：使用 ShareGPT 真实用户数据，效果优于合成数据（真实数据更多样）。\r\n- **数量 vs 质量之争**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e%2BIc4C%2BHdymuGb7PT4VJUhK1USk%3D)\r\n\r\n- **LIMA**：仅用 **1000 条**精选高质量数据微调 65B 模型，效果优于大量数据训练的模型。盲目增加数据量对生成质量无收益。\r\n- **UltraChat**：解决“0-60 分”到“60-100 分”的跨越。定义三大任务维度（信息获取、创造、转化），通过可控设计实现**可扩展的多样性**。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SJcdimmPLr%2FdJDz94jU%2BAuFtnlk%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=L5VbL6nwD45Jge17c5a9kMIRyco%3D)\r\n\r\n- **Orca**：从 Flan 数据集筛选 500 万条输入，用 GPT-4 重写输出。兼顾输入广度与输出质量。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rmhjUoe7sr%2FJL90cljKEQlUJPYY%3D)\r\n\r\n### SFT 的本质理解\r\n1. **0 → 60 分**：学习对话模板与指令遵循，激活预训练知识。\r\n2. **60 → 100 分**：提升推理、生成、知识性及对齐能力，需要更高质量的数据工程。\r\n\r\n\r\n## RLHF —— 基于人类反馈的强化学习\r\n### 为什么需要 RLHF？\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E4uIZWf2gYFdC88dbzVE5nAeVCA%3D)\r\n\r\n- **SFT 的局限**：只能学习“标准答案”，但开放性问题（如写诗、创作）不存在唯一Golden Answer。\r\n- **RLHF 的优势**：通过人类对多个候选答案的**比较与偏好排序**，让模型学习“什么是更好的回答”。\r\n- **长远意义**：从环境\u002F用户反馈中持续学习是智能体的必备能力（适用于多智能体、具身智能等场景）。\r\n\r\n### RLHF 的实践与挑战\r\n1.**OpenAI 的探索**：从 Summarization → WebGPT → InstructGPT。每条 Prompt 生成 4-9 个回复，构建约 20 万条偏好数据；提供详细多维度标注文档辅助人工打分。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fsli1qdZM9Y%2FJ3auQ3LZ5vP1LBE%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QuZtzxkxrK4nxxN9QWtlE%2B%2FTIYQ%3D)\r\n\r\n2.**开源社区的追赶**：\r\n   - **LLaMA 2**：标注 140 万偏好数据，混合开源数据训练奖励模型，成为开源 RLHF 里程碑。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FxIMm%2FvvYyxPEIBx%2FV%2FhVEanKUE%3D)\r\n\r\n   - **痛点**：开源偏好数据少、质量低；小模型做 Reward Model 效果不佳；大模型标数据与顶级专家仍有差距。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=oHc%2FwG6XFr%2BWmZ4%2BOMqSahiS9T8%3D)\r\n\r\n\r\n### UltraFeedback：开源 RLHF 的新标杆\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=N7dyyWruMQhESzCimRtJaccdQUs%3D)\r\n\r\n1. **数据构建**：收集 65K Prompts（来源多样），由 4 个不同模型生成回复以保证多样性。\r\n2. **高质量标注**：使用 GPT-4 从四个维度进行比较打分并给出解释，形成 **38 万条**高质量偏好数据。\r\n3. **正向循环**：UltraFeedback → 训练UltraRM（超越其他开源奖励模型）→ PPO 训练UltraLM → 在多项对话评测中取得了明显提升。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JWENB7hXEOcAMCgMBqNnm8ezmYI%3D)\r\n\r\n\r\n\r\n## 加入训练营\r\n#### 📺完整课程内容，请观看直播或查看课程回放：\r\n**InfiniTensor 官网**： https:\u002F\u002Fwww.infinitensor.com\r\n\r\n**B站直播**： InfiniTensor 官方直播间\r\n\r\n**视频号直播**： InfiniTensor 视频号预约\r\n\r\n**答疑交流**： 训练营官方社群⬇️\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wGzqtB9HsaU4IHS5eDT0qyV2lsg%3D)\r\n\r\n## 关注与交流\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9Am%2FQfw7y%2BpppXGy6g1VG29SISM%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-66","study-66","2026 夏季 InfiniTensor 训练营 CPU 并行编程 1：《CPU 性能优化导论》",42,"\u002Fstudy-material\u002Fstudy-67","study-67","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-68","study-68","2026 夏季 InfiniTensor 训练营 Triton & 九齿 2：《Triton 编译器探索与寒武纪 MLU 实践》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BJteDlJ7AzrQ7Mx%2F%2BOhce5LKx%2Bg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 17 日，2026 夏季 InfiniTensor 训练营前沿模型方向第一课《大模型学习方法》开讲。\u003C\u002Fp>\n\u003Cp>本节课我们将沿着 AI 发展的历史脉络，深入拆解大模型时代的三大核心学习范式：预训练（Pre-training）、监督微调（SFT）以及基于人类反馈的强化学习（RLHF）。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Ch2 id=\"人工智能发展史与核心命题\">\u003Ca class=\"anchor\" href=\"#人工智能发展史与核心命题\" aria-label=\"跳转到标题：人工智能发展史与核心命题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>人工智能发展史与核心命题\u003C\u002Fh2>\u003Ch3 id=\"ai-的核心命题\">\u003Ca class=\"anchor\" href=\"#ai-的核心命题\" aria-label=\"跳转到标题：AI 的核心命题\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>AI 的核心命题\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cp>\u003Cstrong>起源\u003C\u002Fstrong>：1956 年达特茅斯会议标志着 AI 诞生。\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>永恒问题\u003C\u002Fstrong>：怎样让机器掌握完成复杂任务所需的知识？以何种形式存储？如何应用于新任务？\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>AI 发展的三次浪潮与起伏\u003C\u002Fstrong>\u003Cbr>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=eGwYxJdO%2BdQB2gQp3OTHu6KE7JE%3D\" alt=\"\">\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"知识学习的三个代表性阶段\">\u003Ca class=\"anchor\" href=\"#知识学习的三个代表性阶段\" aria-label=\"跳转到标题：知识学习的三个代表性阶段\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>知识学习的三个代表性阶段\u003C\u002Fh3>\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth align=\"left\">阶段\u003C\u002Fth>\n\u003Cth align=\"left\">核心方法\u003C\u002Fth>\n\u003Cth align=\"left\">知识存储形式\u003C\u002Fth>\n\u003Cth align=\"left\">优势\u003C\u002Fth>\n\u003Cth align=\"left\">局限性\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\u003Ctr>\n\u003Ctd align=\"left\">\u003Cstrong>符号智能\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd align=\"left\">专家手写规则\u003C\u002Ftd>\n\u003Ctd align=\"left\">规则库\u002F知识库\u003C\u002Ftd>\n\u003Ctd align=\"left\">清晰、可解释\u003C\u002Ftd>\n\u003Ctd align=\"left\">难以穷举，无泛化能力\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd align=\"left\">\u003Cstrong>专用智能\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd align=\"left\">数据驱动机器学习\u003C\u002Ftd>\n\u003Ctd align=\"left\">小模型参数\u003C\u002Ftd>\n\u003Ctd align=\"left\">自动学习规律，有一定泛化性\u003C\u002Ftd>\n\u003Ctd align=\"left\">依赖标注数据，局限于特定任务\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd align=\"left\">\u003Cstrong>通用智能\u003C\u002Fstrong>\u003C\u002Ftd>\n\u003Ctd align=\"left\">自监督预训练\u003C\u002Ftd>\n\u003Ctd align=\"left\">大模型参数\u003C\u002Ftd>\n\u003Ctd align=\"left\">利用海量无标注数据，支持通用任务\u003C\u002Ftd>\n\u003Ctd align=\"left\">计算效率挑战\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\u003C\u002Ftable>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xA8slZgvzVB%2BIsw6RNBjxsUjn%2Bo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"大模型的特点与能力涌现\">\u003Ca class=\"anchor\" href=\"#大模型的特点与能力涌现\" aria-label=\"跳转到标题：大模型的特点与能力涌现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>大模型的特点与能力涌现\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NqqAW22ie%2F0Lipp3PLkOGbZjzIw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>序列化建模\u003C\u002Fstrong>：文本、图像、音频等多模态信息转化为序列。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>量变到质变\u003C\u002Fstrong>：数据规模↑ → 参数规模↑ → 算力需求↑ → 知识累积 → \u003Cstrong>智能涌现\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>可扩展性\u003C\u002Fstrong>：通过 scaling 获得能力提升，但面临计算效率瓶颈（GPT-3至今参数量级未呈指数级爆炸，受限于效率）。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GVV%2BxLxFd3xld4%2BsjL0J97%2FMMeI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"预训练-pre-training\">\u003Ca class=\"anchor\" href=\"#预训练-pre-training\" aria-label=\"跳转到标题：预训练 （Pre-training）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>预训练 （Pre-training）\u003C\u002Fh2>\u003Ch3 id=\"scaling-law-尺度律\">\u003Ca class=\"anchor\" href=\"#scaling-law-尺度律\" aria-label=\"跳转到标题：Scaling Law（尺度律）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Scaling Law（尺度律）\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ua3cCvoTUu6kX7Znfeq7s61GciY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>核心规律\u003C\u002Fstrong>：在一定范围内，损失函数与参数量、数据量均呈现\u003Cstrong>对数线性下降\u003C\u002Fstrong>（幂次趋势）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>指导意义\u003C\u002Fstrong>：\u003Cul>\n\u003Cli>预测大模型性能，支持训练决策。\u003C\u002Fli>\n\u003Cli>预估给定算力下的最优“参数-数据”配比。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Chinchilla 定律\u003C\u002Fstrong>：参数与数据应以\u003Cstrong>几乎相等的比例\u003C\u002Fstrong>进行扩展，而非单一维度的堆叠。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8dCddeWCa2AOWNo87aydiT2t%2FbM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"4\">\n\u003Cli>\u003Cstrong>关于“涌现”的争议\u003C\u002Fstrong>：阶跃性涌现可能是观测指标的局限，全方位持续观测下能力增长可能是连续的。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"3-数据质量的重要性\">\u003Ca class=\"anchor\" href=\"#3-数据质量的重要性\" aria-label=\"跳转到标题：3. 数据质量的重要性\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 数据质量的重要性\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>Textbooks Are All You Need\u003C\u002Fstrong>：高质量数据可极大改变 Scaling Law 曲线形状。仅用 1\u002F100 数据量的 1.3 B 模型，代码性能可超越 GPT-3.5。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xUi9OjUyyX6El9SNQeXTqZ5cr%2F4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>\u003Cstrong>小模型的潜力\u003C\u002Fstrong>：MiniCPM (2B) 性能超越Llama 13B，接近 Gemma 7B。证明通过优化数据工程和算法，可将更高知识密度压缩进小模型。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qBAcUlFyQ%2B79tpGDyXO2DF6GDd0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"监督微调-sft\">\u003Ca class=\"anchor\" href=\"#监督微调-sft\" aria-label=\"跳转到标题：监督微调（SFT）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>监督微调（SFT）\u003C\u002Fh2>\u003Ch3 id=\"sft-的核心目标\">\u003Ca class=\"anchor\" href=\"#sft-的核心目标\" aria-label=\"跳转到标题：SFT 的核心目标\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>SFT 的核心目标\u003C\u002Fh3>\u003Cul>\n\u003Cli>在广博世界知识基础上，快速强化高级能力。\u003C\u002Fli>\n\u003Cli>约束模型回答偏好，使其符合人类期望。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"指令微调-instruction-tuning\">\u003Ca class=\"anchor\" href=\"#指令微调-instruction-tuning\" aria-label=\"跳转到标题：指令微调（Instruction Tuning）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>指令微调（Instruction Tuning）\u003C\u002Fh3>\u003Cp>1.让模型从“续写者”变为“交互工具”，增强意图理解与泛化能力。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QMdOzzYDZKmYptgelRFaNlY0nX0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.指令微调特性：形式自由\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7xLiWNE0RPYy%2BrZPu%2BAUabU38Wc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>InstructGPT\u003C\u002Fstrong>：引入对话形式微调 + 真实用户指令，是 ChatGPT 的前身。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Alpaca\u003C\u002Fstrong>：Self-Instruct 思路，用大模型构造问答对。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Vicuna\u003C\u002Fstrong>：使用 ShareGPT 真实用户数据，效果优于合成数据（真实数据更多样）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>数量 vs 质量之争\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=e%2BIc4C%2BHdymuGb7PT4VJUhK1USk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>LIMA\u003C\u002Fstrong>：仅用 \u003Cstrong>1000 条\u003C\u002Fstrong>精选高质量数据微调 65B 模型，效果优于大量数据训练的模型。盲目增加数据量对生成质量无收益。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>UltraChat\u003C\u002Fstrong>：解决“0-60 分”到“60-100 分”的跨越。定义三大任务维度（信息获取、创造、转化），通过可控设计实现\u003Cstrong>可扩展的多样性\u003C\u002Fstrong>。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SJcdimmPLr%2FdJDz94jU%2BAuFtnlk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=L5VbL6nwD45Jge17c5a9kMIRyco%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Orca\u003C\u002Fstrong>：从 Flan 数据集筛选 500 万条输入，用 GPT-4 重写输出。兼顾输入广度与输出质量。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rmhjUoe7sr%2FJL90cljKEQlUJPYY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"sft-的本质理解\">\u003Ca class=\"anchor\" href=\"#sft-的本质理解\" aria-label=\"跳转到标题：SFT 的本质理解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>SFT 的本质理解\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>0 → 60 分\u003C\u002Fstrong>：学习对话模板与指令遵循，激活预训练知识。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>60 → 100 分\u003C\u002Fstrong>：提升推理、生成、知识性及对齐能力，需要更高质量的数据工程。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2 id=\"rlhf-基于人类反馈的强化学习\">\u003Ca class=\"anchor\" href=\"#rlhf-基于人类反馈的强化学习\" aria-label=\"跳转到标题：RLHF —— 基于人类反馈的强化学习\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>RLHF —— 基于人类反馈的强化学习\u003C\u002Fh2>\u003Ch3 id=\"为什么需要-rlhf\">\u003Ca class=\"anchor\" href=\"#为什么需要-rlhf\" aria-label=\"跳转到标题：为什么需要 RLHF？\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>为什么需要 RLHF？\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=E4uIZWf2gYFdC88dbzVE5nAeVCA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>SFT 的局限\u003C\u002Fstrong>：只能学习“标准答案”，但开放性问题（如写诗、创作）不存在唯一Golden Answer。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>RLHF 的优势\u003C\u002Fstrong>：通过人类对多个候选答案的\u003Cstrong>比较与偏好排序\u003C\u002Fstrong>，让模型学习“什么是更好的回答”。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>长远意义\u003C\u002Fstrong>：从环境\u002F用户反馈中持续学习是智能体的必备能力（适用于多智能体、具身智能等场景）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"rlhf-的实践与挑战\">\u003Ca class=\"anchor\" href=\"#rlhf-的实践与挑战\" aria-label=\"跳转到标题：RLHF 的实践与挑战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>RLHF 的实践与挑战\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>OpenAI 的探索\u003C\u002Fstrong>：从 Summarization → WebGPT → InstructGPT。每条 Prompt 生成 4-9 个回复，构建约 20 万条偏好数据；提供详细多维度标注文档辅助人工打分。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fsli1qdZM9Y%2FJ3auQ3LZ5vP1LBE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QuZtzxkxrK4nxxN9QWtlE%2B%2FTIYQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>开源社区的追赶\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>LLaMA 2\u003C\u002Fstrong>：标注 140 万偏好数据，混合开源数据训练奖励模型，成为开源 RLHF 里程碑。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FxIMm%2FvvYyxPEIBx%2FV%2FhVEanKUE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>痛点\u003C\u002Fstrong>：开源偏好数据少、质量低；小模型做 Reward Model 效果不佳；大模型标数据与顶级专家仍有差距。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=oHc%2FwG6XFr%2BWmZ4%2BOMqSahiS9T8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"ultrafeedback-开源-rlhf-的新标杆\">\u003Ca class=\"anchor\" href=\"#ultrafeedback-开源-rlhf-的新标杆\" aria-label=\"跳转到标题：UltraFeedback：开源 RLHF 的新标杆\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>UltraFeedback：开源 RLHF 的新标杆\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=N7dyyWruMQhESzCimRtJaccdQUs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>数据构建\u003C\u002Fstrong>：收集 65K Prompts（来源多样），由 4 个不同模型生成回复以保证多样性。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>高质量标注\u003C\u002Fstrong>：使用 GPT-4 从四个维度进行比较打分并给出解释，形成 \u003Cstrong>38 万条\u003C\u002Fstrong>高质量偏好数据。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>正向循环\u003C\u002Fstrong>：UltraFeedback → 训练UltraRM（超越其他开源奖励模型）→ PPO 训练UltraLM → 在多项对话评测中取得了明显提升。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_67_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JWENB7hXEOcAMCgMBqNnm8ezmYI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"加入训练营\">\u003Ca class=\"anchor\" href=\"#加入训练营\" aria-label=\"跳转到标题：加入训练营\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加入训练营\u003C\u002Fh2>\u003Ch4 id=\"完整课程内容-请观看直播或查看课程回放\">\u003Ca class=\"anchor\" href=\"#完整课程内容-请观看直播或查看课程回放\" aria-label=\"跳转到标题：📺完整课程内容，请观看直播或查看课程回放：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📺完整课程内容，请观看直播或查看课程回放：\u003C\u002Fh4>\u003Cp>\u003Cstrong>InfiniTensor 官网\u003C\u002Fstrong>： \u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\">https:\u002F\u002Fwww.infinitensor.com\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Cstrong>B站直播\u003C\u002Fstrong>： InfiniTensor 官方直播间\u003C\u002Fp>\n\u003Cp>\u003Cstrong>视频号直播\u003C\u002Fstrong>： InfiniTensor 视频号预约\u003C\u002Fp>\n\u003Cp>\u003Cstrong>答疑交流\u003C\u002Fstrong>： 训练营官方社群⬇️\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wGzqtB9HsaU4IHS5eDT0qyV2lsg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"关注与交流\">\u003Ca class=\"anchor\" href=\"#关注与交流\" aria-label=\"跳转到标题：关注与交流\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关注与交流\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9Am%2FQfw7y%2BpppXGy6g1VG29SISM%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,133,135,137,140,143,146,149,152,155,158,161,164,167,170,172,176],{"id":129,"text":129,"level":110},"人工智能发展史与核心命题",{"id":131,"text":132,"level":9},"ai-的核心命题","AI 的核心命题",{"id":134,"text":134,"level":9},"知识学习的三个代表性阶段",{"id":136,"text":136,"level":9},"大模型的特点与能力涌现",{"id":138,"text":139,"level":110},"预训练-pre-training","预训练 （Pre-training）",{"id":141,"text":142,"level":9},"scaling-law-尺度律","Scaling Law（尺度律）",{"id":144,"text":145,"level":9},"3-数据质量的重要性","3. 数据质量的重要性",{"id":147,"text":148,"level":110},"监督微调-sft","监督微调（SFT）",{"id":150,"text":151,"level":9},"sft-的核心目标","SFT 的核心目标",{"id":153,"text":154,"level":9},"指令微调-instruction-tuning","指令微调（Instruction Tuning）",{"id":156,"text":157,"level":9},"sft-的本质理解","SFT 的本质理解",{"id":159,"text":160,"level":110},"rlhf-基于人类反馈的强化学习","RLHF —— 基于人类反馈的强化学习",{"id":162,"text":163,"level":9},"为什么需要-rlhf","为什么需要 RLHF？",{"id":165,"text":166,"level":9},"rlhf-的实践与挑战","RLHF 的实践与挑战",{"id":168,"text":169,"level":9},"ultrafeedback-开源-rlhf-的新标杆","UltraFeedback：开源 RLHF 的新标杆",{"id":171,"text":171,"level":110},"加入训练营",{"id":173,"text":174,"level":175},"完整课程内容-请观看直播或查看课程回放","📺完整课程内容，请观看直播或查看课程回放：",4,{"id":177,"text":177,"level":110},"关注与交流"]