feat: launch LLM Atlas research course

2026-07-28 21:55:19 +08:00
commit 7d5c5e4c54
33 changed files with 13285 additions and 0 deletions
@@ -0,0 +1,233 @@
+export type ChapterStatus = "published" | "drafting" | "researching" | "queued";
+
+export interface Chapter {
+  number: string;
+  slug: string;
+  title: string;
+  kicker: string;
+  question: string;
+  summary: string;
+  status: ChapterStatus;
+  progress: number;
+  papers: number;
+  prerequisites: string[];
+  highlights: string[];
+}
+
+export const chapters: Chapter[] = [
+  {
+    number: "00",
+    slug: "roadmap",
+    title: "先看懂一张大模型地图",
+    kicker: "ORIENTATION",
+    question: "面对几百个术语，我们究竟先学什么？",
+    summary: "建立问题链、依赖关系、论文阅读方法和证据等级；把 K3 放回完整技术地图。",
+    status: "published",
+    progress: 72,
+    papers: 12,
+    prerequisites: [],
+    highlights: ["四层难度", "一张依赖图", "阅读路径"],
+  },
+  {
+    number: "01",
+    slug: "foundations/language-models",
+    title: "语言模型从哪里来",
+    kicker: "LANGUAGE MODELING",
+    question: "预测下一个 Token，为什么能产生通用能力？",
+    summary: "从 N-gram、神经概率语言模型、词向量一路走到 Seq2Seq，理解 Transformer 出现前的瓶颈。",
+    status: "researching",
+    progress: 24,
+    papers: 10,
+    prerequisites: [],
+    highlights: ["概率分解", "分布式表示", "序列瓶颈"],
+  },
+  {
+    number: "02",
+    slug: "foundations",
+    title: "注意力与 Transformer",
+    kicker: "TRANSFORMER",
+    question: "一句话里的每个词，怎样直接找到真正相关的词？",
+    summary: "用可操作的小例子拆开 Q、K、V、自注意力、多头、因果掩码、残差与前馈网络。",
+    status: "published",
+    progress: 45,
+    papers: 12,
+    prerequisites: ["01"],
+    highlights: ["Q / K / V", "交互实验", "张量形状"],
+  },
+  {
+    number: "03",
+    slug: "architecture/representation",
+    title: "表示、位置与残差高速公路",
+    kicker: "REPRESENTATION",
+    question: "模型如何知道词序，又如何让信息穿过上百层？",
+    summary: "从分词、位置编码、归一化与激活函数，走到深层网络的信息流和 Attention Residuals。",
+    status: "researching",
+    progress: 18,
+    papers: 16,
+    prerequisites: ["02"],
+    highlights: ["RoPE", "RMSNorm", "AttnRes"],
+  },
+  {
+    number: "04",
+    slug: "scaling",
+    title: "Scaling Laws：规模为什么有效",
+    kicker: "SCALING",
+    question: "多大模型、多少数据、多少计算才是划算的？",
+    summary: "区分参数量、激活参数、训练计算和推理预算；理解 Kaplan、Chinchilla 与 K3 的双轴扩展。",
+    status: "researching",
+    progress: 22,
+    papers: 11,
+    prerequisites: ["01", "02"],
+    highlights: ["幂律", "计算最优", "测试时扩展"],
+  },
+  {
+    number: "05",
+    slug: "pretraining/data",
+    title: "数据工程与预训练配方",
+    kicker: "DATA",
+    question: "更多网页为什么不等于更好的模型？",
+    summary: "追踪采集、过滤、去重、混合、课程、合成数据与污染控制，并明确公开报告的知识边界。",
+    status: "queued",
+    progress: 8,
+    papers: 14,
+    prerequisites: ["01", "04"],
+    highlights: ["数据质量", "混合策略", "污染"],
+  },
+  {
+    number: "06",
+    slug: "architecture/moe",
+    title: "稀疏计算与 MoE",
+    kicker: "SPARSE EXPERTS",
+    question: "怎样让模型装下更多知识，却不让每个 Token 都付全部算力？",
+    summary: "从条件计算到 DeepSeekMoE、LatentMoE 与 K3 Stable LatentMoE，解释路由、特化和负载均衡。",
+    status: "researching",
+    progress: 28,
+    papers: 17,
+    prerequisites: ["02", "04"],
+    highlights: ["专家路由", "细粒度专家", "896 选 16"],
+  },
+  {
+    number: "07",
+    slug: "architecture/long-context",
+    title: "长上下文与高效注意力",
+    kicker: "LONG CONTEXT",
+    question: "从 8K 到 1M Token，真正昂贵的是什么？",
+    summary: "比较稀疏/线性注意力、FlashAttention、MLA、状态空间模型、Delta Rule、KDA 与混合注意力。",
+    status: "researching",
+    progress: 31,
+    papers: 24,
+    prerequisites: ["02", "03"],
+    highlights: ["KV Cache", "MLA", "KDA"],
+  },
+  {
+    number: "08",
+    slug: "systems/training",
+    title: "大规模训练系统",
+    kicker: "DISTRIBUTED TRAINING",
+    question: "一个 2.8T 模型如何摊到成千上万张卡上？",
+    summary: "从 ZeRO、Megatron 和五类并行，走到 DeepSeek DualPipe/DeepEP 与 K3 MoonEP。",
+    status: "queued",
+    progress: 12,
+    papers: 18,
+    prerequisites: ["04", "06"],
+    highlights: ["并行维度", "通信重叠", "MoonEP"],
+  },
+  {
+    number: "09",
+    slug: "systems/numerics",
+    title: "数值精度、优化器与稳定性",
+    kicker: "OPTIMIZATION",
+    question: "为什么少几个比特能省巨资，也可能让训练瞬间崩掉？",
+    summary: "解释 AdamW、μP、Muon、BF16/FP8/MXFP4 与量化感知训练的数值直觉。",
+    status: "queued",
+    progress: 13,
+    papers: 16,
+    prerequisites: ["02", "08"],
+    highlights: ["FP8", "Muon", "MXFP4"],
+  },
+  {
+    number: "10",
+    slug: "post-training/alignment",
+    title: "指令微调与人类偏好",
+    kicker: "ALIGNMENT",
+    question: "会续写的 base model，怎样变成愿意协作的助手？",
+    summary: "从 instruction tuning、SFT、RLHF/PPO 到 DPO 与 RLAIF，区分能力学习和行为塑形。",
+    status: "queued",
+    progress: 10,
+    papers: 18,
+    prerequisites: ["01", "02"],
+    highlights: ["SFT", "奖励模型", "偏好优化"],
+  },
+  {
+    number: "11",
+    slug: "reasoning",
+    title: "推理模型与测试时扩展",
+    kicker: "REASONING",
+    question: "模型如何学会多想一会儿，并检查自己的答案？",
+    summary: "从 CoT、搜索与验证器，到 GRPO、DeepSeek-R1、Kimi k1.5 和 multi-effort RL。",
+    status: "researching",
+    progress: 25,
+    papers: 21,
+    prerequisites: ["10"],
+    highlights: ["GRPO", "R1-Zero", "On-policy 蒸馏"],
+  },
+  {
+    number: "12",
+    slug: "agents",
+    title: "工具使用与长程 Agent",
+    kicker: "AGENTS",
+    question: "生成一段文字，怎样变成持续数小时的可靠行动？",
+    summary: "讨论工具协议、行动—观察循环、环境奖励、沙箱、可验证任务与百万 Token 轨迹。",
+    status: "queued",
+    progress: 14,
+    papers: 22,
+    prerequisites: ["10", "11"],
+    highlights: ["ReAct", "沙箱", "长轨迹 RL"],
+  },
+  {
+    number: "13",
+    slug: "multimodal",
+    title: "原生多模态",
+    kicker: "MULTIMODAL",
+    question: "图像是外接插件，还是和文字一样的第一类输入？",
+    summary: "从 ViT/CLIP、连接器式 VLM 走到 Kimi-VL、MoonViT-V2 与统一主干。",
+    status: "queued",
+    progress: 11,
+    papers: 19,
+    prerequisites: ["02"],
+    highlights: ["ViT", "视觉 Token", "MoonViT-V2"],
+  },
+  {
+    number: "14",
+    slug: "systems/inference",
+    title: "推理服务与低成本部署",
+    kicker: "INFERENCE",
+    question: "模型训练完以后，怎样让千万人用得起？",
+    summary: "拆解 KV Cache、PagedAttention、批处理、推测解码、PD 解耦、前缀缓存与集群调度。",
+    status: "queued",
+    progress: 12,
+    papers: 20,
+    prerequisites: ["07", "08", "09"],
+    highlights: ["vLLM", "Mooncake", "KDA 缓存"],
+  },
+  {
+    number: "15",
+    slug: "evaluation",
+    title: "评测、安全与“到底强不强”",
+    kicker: "EVALUATION",
+    question: "一个榜单分数，究竟测到了模型、脚手架还是预算？",
+    summary: "从语言建模指标走到知识、代码、Agent 和多模态评测，识别污染、harness 与选择性报告。",
+    status: "queued",
+    progress: 13,
+    papers: 24,
+    prerequisites: ["04", "10", "12"],
+    highlights: ["基准演化", "Harness", "安全边界"],
+  },
+];
+
+export const statusLabel: Record<ChapterStatus, string> = {
+  published: "首版可读",
+  drafting: "写作中",
+  researching: "研究中",
+  queued: "待展开",
+};