feat: publish data engineering chapter
This commit is contained in:
+17
-3
@@ -11,6 +11,7 @@
|
||||
| Kimi K3 深读 | 进行中 | 66% | 扩写 pre-training / infra 逐图笔记 |
|
||||
| Transformer 基础 | 进行中 | 52% | 矩阵形状动画与手算练习 |
|
||||
| Scaling Laws | 完成首版 | 74% | 真实拟合复现、置信区间与更多模型族对照 |
|
||||
| 数据工程与预训练配方 | 完成首版 | 73% | FineWeb / DCLM 逐图精读、真实去重误伤与 mixture traces |
|
||||
| DeepSeek 专题 | 进行中 | 71% | 补 R1 / DAPO 的逐图训练轨迹与复现对照 |
|
||||
| 推理与测试时扩展 | 完成首版 | 76% | 真实模型采样曲线、PRM 案例与逐篇图表精读 |
|
||||
| 稀疏计算与 MoE | 完成首版 | 74% | 真实负载 traces 与专家特化案例 |
|
||||
@@ -29,10 +30,10 @@
|
||||
- [x] 提炼参考网站的编辑设计语言。
|
||||
- [x] 确认 `git.k1412.top` 为 Gitea/Forgejo 兼容服务且本机 HTTPS 凭据可用于既有仓库。
|
||||
- [x] 使用 Grok CLI 检索并形成约 95 篇一手论文的补充路线,主代理已回查关键来源。
|
||||
- [x] 完成 182 篇关键论文索引,覆盖 12 个专题与 Kimi/DeepSeek 聚光主线。
|
||||
- [x] 完成 202 篇关键论文索引,覆盖 13 个标签专题与 Kimi/DeepSeek 聚光主线。
|
||||
- [x] 完成可检索、可按专题筛选的论文库页面。
|
||||
- [x] 完成 K3、Transformer 基础、DeepSeek 谱系、Scaling Laws、长上下文、MoE、推理与训练系统八篇首版长文。
|
||||
- [x] 完成 K3 三轴架构、Self-Attention、DeepSeek 谱系、长上下文、MoE 路由、推理三页签,以及训练系统与 Scaling 各四页签等十六个原创交互视图。
|
||||
- [x] 完成 K3、Transformer 基础、DeepSeek 谱系、Scaling Laws、数据工程、长上下文、MoE、推理与训练系统九篇首版长文。
|
||||
- [x] 完成 K3 三轴架构、Self-Attention、DeepSeek 谱系、长上下文、MoE 路由、推理三页签,以及训练系统、Scaling 与数据工程各四页签等二十个原创交互视图。
|
||||
- [x] 完成长上下文首版:五张成本账、26 篇一手论文、10+ 机制图与 8 策略交互实验室。
|
||||
- [x] 核验 FlashAttention、DeepSeek-V2/V3.2/V4、Kimi Linear/K3 等六份论文原文,并建立长上下文研究账本。
|
||||
- [x] 核验 Switch、ST-MoE、DeepSeekMoE、Loss-Free、V3、LatentMoE 与 K3 原文,并建立 MoE 研究账本。
|
||||
@@ -58,10 +59,18 @@
|
||||
- [x] Scaling 真实 Chrome 断言通过:IsoFLOP 谷底、生命周期最优点、重复数据衰减和指标阶跃均正确响应。
|
||||
- [x] 对 Scaling、训练系统、推理与 MoE 执行全量浏览器回归;12 项桌面导航、移动菜单和首页新章入口无回归。
|
||||
- [x] Scaling 首版以提交 `eff0e4c`、不可变镜像 `20260728T181742Z-eff0e4c` 发布;NAS、VPS/Tailscale、NPM、DNS、HTTPS、证书、12 路由与生产 Chrome 全链路通过。
|
||||
- [x] 启动数据工程专题:回查 K3/K2、DeepSeek LLM/Math/V2/V3/V4 数据章节,并将 Grok 线索永久隔离为未核验候选。
|
||||
- [x] 建立 1264 行正式研究账本:十二张账、2019→2026 历史主线、DeepSeek/Kimi 双谱系、四实验合同与公开未知项。
|
||||
- [x] 完成数据工程首版:18 节长文、31 个一手节点、十余组机制图与流水线—去重—混合—改写四联实验。
|
||||
- [x] 论文库新增「数据」标签与 20 个一手节点,从 182 篇扩充至 202 篇。
|
||||
- [x] 数据工程真实 Chrome 断言通过:激进过滤、semantic 去重误伤、数学 mixture 和失控合成均正确响应,桌面 / 390px 移动端无溢出。
|
||||
- [x] 对 Scaling、训练系统、推理与 MoE 重新执行全量浏览器回归;13 项桌面 / 移动导航和首页数据新章入口无回归。
|
||||
- [x] Astro 类型检查、生产构建、13 个页面路由、527 个站内引用和 20 个跨页锚点全部通过。
|
||||
|
||||
## 正在进行
|
||||
|
||||
- [ ] Scaling Laws 二轮精读:真实拟合复现、逐篇图表、置信区间、外推失败与更多模型族对照。
|
||||
- [ ] 数据工程二轮:FineWeb / DCLM / Dolma / ROOTS 逐图精读、真实去重误伤、mixture traces 与污染案例。
|
||||
- [ ] 大规模训练系统二轮:真实集群 traces、故障恢复案例与精确 topology / kernel 配置。
|
||||
- [ ] 推理专题二轮:真实 pass@k 曲线、PRM 失败案例与逐篇图表精读。
|
||||
- [ ] 长上下文专题的真实模型配置对比、内核细节与失败案例二轮深化。
|
||||
@@ -98,6 +107,11 @@
|
||||
| 2026-07-29 | Scaling 首版用 29 个一手节点和四个独立实验闭环 | IsoFLOP、训练—部署、数据复用与指标涌现分开建模,不合成伪“规模总分” |
|
||||
| 2026-07-29 | 论文库随 Scaling 一手链扩充到 182 篇 | 新增早期跨任务幂律、Gopher、数据受限、过训练、部署经济、复现和 task ladder 节点 |
|
||||
| 2026-07-29 | Scaling 首版用不可变镜像 `20260728T181742Z-eff0e4c` 发布 | OCI digest `sha256:802e4041…cf83`;保留 `12010→8080`、NPM host 31 / cert 41、门户 order 180 与公开 Forgejo |
|
||||
| 2026-07-29 | 数据工程拆成来源、解析、语言、质量、唯一性、污染、混合、变换、Tokenizer、Packing、课程与价值十二张账 | 防止用一个“高质量 Token”标签掩盖单位、误删、治理与课程 |
|
||||
| 2026-07-29 | Grok 数据研究包只保留为 `DATA_PRETRAINING_GROK_LEADS.md` | 模型发现候选不越过一手证据门槛;K3 总训练 Token、工业 mixture 和 source map 保持未知 |
|
||||
| 2026-07-29 | K2 改写、K3 2.5×、V3/V4 packing 边界在正文主视区显式限定 | 不把早期 SimpleQA 对照外推为普遍定律,不把模型族收益归因给单一数据技巧 |
|
||||
| 2026-07-29 | 数据工程首版用四个独立实验闭环 | 文档保留、去重误伤、领域 exposure 与改写课程分开建模,不合成伪“数据质量总分” |
|
||||
| 2026-07-29 | 论文库新增数据标签并扩充到 202 篇 | 新增 CCNet、ROOTS、Dedup、DoReMi、DCLM、WRAP、MATES、MM1、Swallow 等 20 个节点 |
|
||||
|
||||
## 未决问题
|
||||
|
||||
|
||||
@@ -17,8 +17,8 @@
|
||||
- 持续进度:[PROGRESS.md](./PROGRESS.md)
|
||||
- 证据与写作规范:[research/METHODOLOGY.md](./research/METHODOLOGY.md)
|
||||
|
||||
当前里程碑包含 16 专题学习地图、182 篇关键论文索引、Kimi K3 完整导读、
|
||||
Transformer 基础、DeepSeek 技术谱系、Scaling Laws、长上下文、MoE、推理与训练系统深度专题,
|
||||
当前里程碑包含 16 专题学习地图、202 篇关键论文索引、Kimi K3 完整导读、
|
||||
Transformer 基础、DeepSeek 技术谱系、Scaling Laws、数据工程、长上下文、MoE、推理与训练系统深度专题,
|
||||
以及覆盖核心机制的原创交互实验室。
|
||||
其余专题按进度账本持续扩建。
|
||||
|
||||
|
||||
+2
-1
@@ -13,7 +13,8 @@
|
||||
"check:moe-browser": "node scripts/check-moe-browser.mjs",
|
||||
"check:reasoning-browser": "node scripts/check-reasoning-browser.mjs",
|
||||
"check:training-systems-browser": "node scripts/check-training-systems-browser.mjs",
|
||||
"check:scaling-browser": "node scripts/check-scaling-browser.mjs"
|
||||
"check:scaling-browser": "node scripts/check-scaling-browser.mjs",
|
||||
"check:data-browser": "node scripts/check-data-browser.mjs"
|
||||
},
|
||||
"dependencies": {
|
||||
"@astrojs/sitemap": "3.7.3",
|
||||
|
||||
@@ -0,0 +1,203 @@
|
||||
# 数据工程与预训练配方 · Grok Discovery Leads
|
||||
|
||||
> **未核验线索,不是正式研究账本。**
|
||||
>
|
||||
> 2026-07-29 使用本机 Grok CLI Headless 扩展候选论文、关键词与视觉方案。
|
||||
> 正式结论只进入 `DATA_PRETRAINING_RESEARCH.md`,且必须回到论文、技术报告、数据卡或作者实现。
|
||||
> Grok 被明确禁止调用子代理;其输出只承担“还有什么值得查”,不承担“事实究竟是什么”。
|
||||
|
||||
---
|
||||
|
||||
## 1. Grok 建议的十二张候选账
|
||||
|
||||
| 账 | 候选问题 | 候选关键词 |
|
||||
|---|---|---|
|
||||
| L1 来源与治理 | 数据来自哪里,是否允许获取、研究、训练与再分发 | data card、license、consent、robots、provenance、governance |
|
||||
| L2 提取与解析 | HTML、PDF、代码与表格怎样变成没有被破坏的文档 | boilerplate removal、DOM、OCR、code parser、structural validation |
|
||||
| L3 语言与领域 | 语言识别、领域识别和长尾覆盖怎样度量 | fastText、language ID、domain classifier、long tail |
|
||||
| L4 质量过滤 | “高质量”由哪些可观测量定义 | heuristic、classifier、educational value、toxicity、PII |
|
||||
| L5 去重 | exact、fuzzy、semantic 各去掉什么,又会误删什么 | hash、MinHash、LSH、embedding、cluster、frequency |
|
||||
| L6 混合与采样 | 数据占比怎样变成模型能力的组合 | mixture weight、temperature、upsampling、proxy model、DoReMi |
|
||||
| L7 Tokenizer 与 Packing | 文档怎样变成 Token 和训练序列 | BBPE、compression、BPB、best-fit packing、sample mask、FIM |
|
||||
| L8 课程与退火 | 不同数据和长度何时进入训练 | curriculum、annealing、mid-training、continual pretraining、WSD |
|
||||
| L9 合成与改写 | 重复、改写、翻译和生成数据有何不同 | WRAP、rephrasing、fidelity、teacher bias、model collapse |
|
||||
| L10 污染与记忆 | benchmark、隐私和长串复现如何被发现 | decontamination、memorization、extraction、canary |
|
||||
| L11 原生多模态 | 图片、视频、OCR、代码渲染怎样与文本共同训练 | caption、interleaved、MM1、native multimodal、coordinate supervision |
|
||||
| L12 数据价值 | 哪个样本值得下一个训练 FLOP | influence、gradient matching、DSIR、MATES、JEST、data selection |
|
||||
|
||||
这十二张账是 discovery taxonomy,不表示这些问题可以被压缩成一个“数据质量总分”。
|
||||
|
||||
---
|
||||
|
||||
## 2. 候选历史节点
|
||||
|
||||
### 2019–2021:把互联网变成可训练语料
|
||||
|
||||
- T5 / C4:Common Crawl 清洗与 text-to-text 统一任务;
|
||||
- CCNet:去重、语言识别与近似 Wikipedia 质量过滤;
|
||||
- The Pile:22 个异质子语料与明确 mixture;
|
||||
- ROOTS / BigScience:多语言、治理、数据卡与协作式语料建设;
|
||||
- Gopher:大规模网页质量分类与 mixture 工程;
|
||||
- Carlini et al.:训练数据抽取与隐私记忆风险。
|
||||
|
||||
### 2022–2023:去重、混合与数据中心化
|
||||
|
||||
- Lee et al.:近重复会污染验证集并放大记忆;
|
||||
- FIM:预训练数据变换可以增加代码“补中间”能力;
|
||||
- DoReMi:用小代理模型学习 domain weights;
|
||||
- RefinedWeb:过滤网页数据可以支撑强模型;
|
||||
- SemDeDup:embedding 空间中的语义去重;
|
||||
- D4:去重之后继续做多样性选择;
|
||||
- DSIR:按目标分布重要性重采样。
|
||||
|
||||
### 2024:公开数据工程成为可比较实验
|
||||
|
||||
- Dolma:3T English Token、开放工具与文档;
|
||||
- FineWeb:大规模网页数据的清洗消融;
|
||||
- DCLM:统一 pool、模型尺度与评测协议;
|
||||
- WRAP:改写网页文档而不是只做原文多 epoch;
|
||||
- MATES:模型感知的数据选择;
|
||||
- SoftDedup:按频率降权,而非“一刀切”删除;
|
||||
- Best-fit Packing:减少文档截断;
|
||||
- MM1:多模态数据 mixture 消融;
|
||||
- DeepSeek LLM / Math / V2 / V3:全局去重、领域构造、packing、FIM 与训练配方。
|
||||
|
||||
### 2025–2026:数据稀缺、合成扩展与超长上下文
|
||||
|
||||
- Kimi K2:知识与数学改写、15.5T Token、WSD 与长上下文激活;
|
||||
- Swallow:代码与数学合成数据的受控持续预训练;
|
||||
- Common Pile:更强调公开许可的数据源组合;
|
||||
- Group-MATES:group-level data influence;
|
||||
- DeepSeek-V4:过滤批量自动生成 / 模板内容、长文档与 agentic mid-training;
|
||||
- Kimi K3:四文本域 + 大规模视觉语料、原生多模态与 8K → 64K → 256K → 1M 课程。
|
||||
|
||||
---
|
||||
|
||||
## 3. 候选论文池
|
||||
|
||||
下面只表示“值得回查”。表中的数字、结论和适用范围尚未因出现在这里而获得证据等级。
|
||||
|
||||
| 方向 | 候选来源 |
|
||||
|---|---|
|
||||
| 网页清洗 | T5/C4、CCNet、RefinedWeb、FineWeb、DCLM |
|
||||
| 开放语料 | The Pile、ROOTS、Dolma、Common Pile |
|
||||
| 质量分类 | Gopher data pipeline、FineWeb-Edu、DCLM fastText classifier |
|
||||
| exact / fuzzy dedup | Lee et al. 2021、SlimPajama、Dolma |
|
||||
| semantic dedup | SemDeDup、D4、SoftDedup |
|
||||
| mixture | DoReMi、DSIR、Data Mixing Laws、RegMix |
|
||||
| packing | Fewer Truncations Improve Language Modeling |
|
||||
| FIM | Efficient Training of Language Models to Fill in the Middle |
|
||||
| 改写 | WRAP、Swallow、Kimi K2、Kimi K3 |
|
||||
| 动态数据价值 | MATES、Group-MATES、LESS、JEST |
|
||||
| 污染 | Carlini extraction、Detecting Pretraining Data from LLMs、Contamination Detection |
|
||||
| 多模态 mixture | MM1、JEST、Kimi K2.5、Kimi K3 |
|
||||
| 长上下文数据 | DeepSeek-V4、Kimi K3、Long-context synthetic retrieval |
|
||||
|
||||
---
|
||||
|
||||
## 4. 候选视觉
|
||||
|
||||
1. **一条文档的生命线**
|
||||
来源 → 下载 → 解析 → 语言 / 领域 → 质量 / 安全 → 去重 → 污染隔离 → 混合 → 改写 / FIM → Tokenize → Packing → Batch。
|
||||
|
||||
2. **去重显微镜**
|
||||
同时展示 exact hash、shingle MinHash 与 embedding 相似度;拖动阈值观察漏删、误删和长尾损失。
|
||||
|
||||
3. **Mixture Studio**
|
||||
调整 Web / Code / Math / Knowledge / Multimodal 权重;分开展示通用、代码、数学、多语言和平均 loss,不制造“万能能力分”。
|
||||
|
||||
4. **改写与重复实验**
|
||||
对比 raw × epochs、faithful rephrase、错误 rephrase;展示 token utility、事实保真、风格多样性和生成成本。
|
||||
|
||||
5. **课程时间轴**
|
||||
Token 轴上叠加学习率、batch、数据 mixture、context length、FIM 率与 synthetic 比例。
|
||||
|
||||
6. **污染火灾图**
|
||||
benchmark 文档进入训练语料后,按 exact、n-gram、semantic、ground-truth overlap 不同检查器显示不同告警。
|
||||
|
||||
---
|
||||
|
||||
## 5. 候选互动实验
|
||||
|
||||
### 实验 A:Pipeline Ledger
|
||||
|
||||
- 输入:抓取量、解析通过率、质量阈值、去重阈值、许可 / PII 隔离策略;
|
||||
- 输出:文档数、唯一 Token、各领域保留率、未知来源比例;
|
||||
- 必须展示:减少的“量”不自动等于增加的“质量”。
|
||||
|
||||
### 实验 B:Dedup Lens
|
||||
|
||||
- 输入:exact / fuzzy / semantic 模式、相似度阈值、保留策略;
|
||||
- 输出:duplicate precision / recall、benchmark overlap、长尾簇被删比例;
|
||||
- 必须展示:同一句模板与真正重复文档不是同一个问题。
|
||||
|
||||
### 实验 C:Mixture Studio
|
||||
|
||||
- 输入:领域权重、temperature、目标验证集;
|
||||
- 输出:各领域 loss / score 的 Pareto 变化;
|
||||
- 必须展示:最优 mixture 依赖模型尺度、训练阶段与目标分布。
|
||||
|
||||
### 实验 D:Rewrite & Curriculum
|
||||
|
||||
- 输入:改写次数、fidelity、重复 epoch、context stage、长文 upsampling;
|
||||
- 输出:有效新颖度、错误扩增、训练成本、长上下文覆盖;
|
||||
- 必须展示:synthetic token 不是无成本的新事实。
|
||||
|
||||
---
|
||||
|
||||
## 6. 必须回到原文的高风险数字
|
||||
|
||||
- 任意语料的“原始文档数”“最终 Token 数”和过滤保留率;
|
||||
- dedup 删除百分比,以及它是按文档、字节还是 Token 计;
|
||||
- DCLM、FineWeb、WRAP、MATES、SoftDedup 的作者报告收益;
|
||||
- Kimi K2 的 SimpleQA 改写消融;
|
||||
- Kimi K2、DeepSeek-V3 / V4 的训练 Token 与 schedule;
|
||||
- DeepSeek LLM 91 个 Common Crawl dump 的去重率;
|
||||
- DeepSeekMath 的 120.2B corpus 和 500B continual-pretraining mixture;
|
||||
- K3 的四文本域、视觉 taxonomy、原生多模态与长上下文阶段;
|
||||
- 任何“X 倍更高效”的分母、模型大小、任务和预算。
|
||||
|
||||
---
|
||||
|
||||
## 7. 已在主代理回查中纠正的风险
|
||||
|
||||
1. **K3 没有披露总训练 Token 数。**
|
||||
`2.78T` 是模型总参数,不是训练 Token;正式报告只描述数据域、pipeline 和训练阶段。
|
||||
|
||||
2. **K3 的 `2.5×` 不是某个数据技巧的独立收益。**
|
||||
报告 §3.2 明确说它来自架构、数据和训练变化共同定义的新模型族,并在 held-out OOD scaling curve 上比较。
|
||||
|
||||
3. **K2 改写表不是普遍定律。**
|
||||
`23.76 → 27.39 → 28.94` 来自 K2 早期 checkpoint、SimpleQA 与三种指定配置;生产中每个 corpus 最多改写两次。
|
||||
|
||||
4. **K2 的 WSD 与 K3 的 cosine 不构成简单胜负。**
|
||||
K3 分别为两种 schedule 独立搜索 peak LR 和 batch 后才报告 cosine 更低;共享超参数会造成不公平比较。
|
||||
|
||||
5. **DeepSeek-V3 与 V4 的 packing 边界不同。**
|
||||
V3 做 document packing 且不使用 cross-sample attention mask;V4 沿用 packing,但改为 sample-level attention masking。
|
||||
|
||||
6. **DeepSeek 的工业级语料仍是部分黑箱。**
|
||||
报告公开阶段、部分总量和消融,但没有完整 source-by-source license map、最终 mixture、过滤阈值与 contamination 清单。
|
||||
|
||||
7. **“高质量”不能用 benchmark 提升循环定义。**
|
||||
必须分别记录分类器目标、人工标注协议、保留率、误删率、长尾覆盖与独立验证。
|
||||
|
||||
8. **合成数据会继承生成模型偏差。**
|
||||
fidelity filter 不能证明事实完全正确,也不能消除风格坍缩、毒性、版权、隐私或污染风险。
|
||||
|
||||
---
|
||||
|
||||
## 8. 从 leads 晋级正式账本的门槛
|
||||
|
||||
候选结论至少满足:
|
||||
|
||||
1. 找到稳定的一手入口;
|
||||
2. 核对标题、作者、版本与年份;
|
||||
3. 读到支持结论的正文 / 表格 / 附录,而不仅是搜索摘要;
|
||||
4. 记录实验对象、模型大小、Token 预算、数据与评测协议;
|
||||
5. 区分作者报告、独立复现与课程推导;
|
||||
6. 明确失败条件与不可外推范围;
|
||||
7. 数字能追到表、图、公式或可复现实验;
|
||||
8. 无法公开确认的工业细节明确写为“未知”。
|
||||
|
||||
这份文件的价值不是“收集得多”,而是把发现阶段与证据阶段永久隔开。
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,249 @@
|
||||
import { writeFileSync } from "node:fs";
|
||||
|
||||
const cdpPort = process.env.CDP_PORT ?? "9226";
|
||||
const baseUrl = process.env.SITE_URL ?? "http://127.0.0.1:4325";
|
||||
const pages = await fetch(`http://127.0.0.1:${cdpPort}/json/list`).then((response) => response.json());
|
||||
const page = pages.find((entry) => entry.type === "page");
|
||||
if (!page) throw new Error(`CDP ${cdpPort} 没有可用页面`);
|
||||
|
||||
const socket = new WebSocket(page.webSocketDebuggerUrl);
|
||||
await new Promise((resolve, reject) => {
|
||||
socket.addEventListener("open", resolve, { once: true });
|
||||
socket.addEventListener("error", reject, { once: true });
|
||||
});
|
||||
|
||||
let nextId = 0;
|
||||
const pending = new Map();
|
||||
const exceptions = [];
|
||||
socket.addEventListener("message", (event) => {
|
||||
const message = JSON.parse(event.data);
|
||||
if (message.id && pending.has(message.id)) {
|
||||
const { resolve, reject } = pending.get(message.id);
|
||||
pending.delete(message.id);
|
||||
if (message.error) reject(new Error(message.error.message));
|
||||
else resolve(message.result);
|
||||
}
|
||||
if (message.method === "Runtime.exceptionThrown") {
|
||||
exceptions.push(message.params.exceptionDetails.exception?.description ?? message.params.exceptionDetails.text);
|
||||
}
|
||||
});
|
||||
|
||||
const command = (method, params = {}) => new Promise((resolve, reject) => {
|
||||
const id = ++nextId;
|
||||
pending.set(id, { resolve, reject });
|
||||
socket.send(JSON.stringify({ id, method, params }));
|
||||
});
|
||||
const pause = (milliseconds) => new Promise((resolve) => setTimeout(resolve, milliseconds));
|
||||
const evaluate = async (expression) => {
|
||||
const result = await command("Runtime.evaluate", { expression, returnByValue: true, awaitPromise: true });
|
||||
if (result.exceptionDetails) throw new Error(result.exceptionDetails.exception?.description ?? result.exceptionDetails.text);
|
||||
return result.result.value;
|
||||
};
|
||||
const navigate = async (path) => {
|
||||
await command("Page.navigate", { url: `${baseUrl}${path}` });
|
||||
for (let attempt = 0; attempt < 50; attempt += 1) {
|
||||
await pause(100);
|
||||
if (await evaluate("document.readyState === 'complete'")) return;
|
||||
}
|
||||
throw new Error(`${path} 加载超时`);
|
||||
};
|
||||
const screenshot = async (path) => {
|
||||
const result = await command("Page.captureScreenshot", { format: "png", captureBeyondViewport: false });
|
||||
writeFileSync(path, Buffer.from(result.data, "base64"));
|
||||
};
|
||||
|
||||
await command("Page.enable");
|
||||
await command("Runtime.enable");
|
||||
await command("Emulation.setDeviceMetricsOverride", {
|
||||
width: 1440,
|
||||
height: 1100,
|
||||
deviceScaleFactor: 1,
|
||||
mobile: false,
|
||||
});
|
||||
await navigate("/pretraining/data/");
|
||||
await screenshot("/tmp/llm-atlas-data-desktop.png");
|
||||
|
||||
const pipeline = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-data-lab]");
|
||||
const measure = () => ({
|
||||
trainable: root.querySelector('[data-pipeline-metric="trainable"]').textContent,
|
||||
retention: root.querySelector('[data-pipeline-metric="retention"]').textContent,
|
||||
quarantine: root.querySelector('[data-pipeline-metric="quarantine"]').textContent,
|
||||
rows: root.querySelectorAll(".funnel-row").length,
|
||||
tail: root.querySelector('[data-pipeline-slice-label="tail"]').textContent,
|
||||
});
|
||||
const balanced = measure();
|
||||
root.querySelector('[data-pipeline-preset="strict"]').click();
|
||||
return { balanced, strict: measure() };
|
||||
})()`);
|
||||
|
||||
const dedup = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-data-lab]");
|
||||
root.querySelector('[data-data-tab="dedup"]').click();
|
||||
const measure = () => ({
|
||||
precision: root.querySelector('[data-dedup-metric="precision"]').textContent,
|
||||
recall: root.querySelector('[data-dedup-metric="recall"]').textContent,
|
||||
falsePositive: root.querySelector('[data-dedup-metric="false"]').textContent,
|
||||
rare: root.querySelector('[data-dedup-metric="rare"]').textContent,
|
||||
rows: root.querySelectorAll("[data-dedup-example]").length,
|
||||
});
|
||||
root.querySelector('[data-dedup-mode="exact"]').click();
|
||||
const exact = measure();
|
||||
root.querySelector('[data-dedup-mode="semantic"]').click();
|
||||
const threshold = root.querySelector('[data-dedup-input="threshold"]');
|
||||
threshold.value = "85";
|
||||
threshold.dispatchEvent(new Event("input", { bubbles: true }));
|
||||
return { exact, semantic: measure() };
|
||||
})()`);
|
||||
|
||||
const mixture = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-data-lab]");
|
||||
root.querySelector('[data-data-tab="mixture"]').click();
|
||||
const measure = () => ({
|
||||
weights: [...root.querySelectorAll("[data-mixture-output]")].slice(0, 5).map((item) => item.textContent),
|
||||
epochs: [...root.querySelectorAll("[data-mixture-ledger] small")].map((item) => item.textContent),
|
||||
scores: [...root.querySelectorAll("[data-mixture-slices] strong")].map((item) => item.textContent),
|
||||
rows: root.querySelectorAll("[data-mixture-ledger] > div").length,
|
||||
});
|
||||
root.querySelector('[data-mixture-preset="balanced"]').click();
|
||||
const balanced = measure();
|
||||
root.querySelector('[data-mixture-preset="math"]').click();
|
||||
return { balanced, math: measure() };
|
||||
})()`);
|
||||
|
||||
const transform = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-data-lab]");
|
||||
root.querySelector('[data-data-tab="transform"]').click();
|
||||
const measure = () => ({
|
||||
variants: root.querySelector('[data-transform-metric="variants"]').textContent,
|
||||
errors: root.querySelector('[data-transform-metric="errors"]').textContent,
|
||||
cost: root.querySelector('[data-transform-metric="cost"]').textContent,
|
||||
context: root.querySelector('[data-transform-metric="context"]').textContent,
|
||||
stages: root.querySelectorAll("[data-transform-timeline] > div").length,
|
||||
errorDots: root.querySelectorAll('[data-transform-facts] > i[data-kind="error"]').length,
|
||||
});
|
||||
root.querySelector('[data-transform-preset="production"]').click();
|
||||
const production = measure();
|
||||
root.querySelector('[data-transform-preset="risky"]').click();
|
||||
const risky = measure();
|
||||
const firstTab = root.querySelector('[data-data-tab="pipeline"]');
|
||||
firstTab.focus();
|
||||
firstTab.dispatchEvent(new KeyboardEvent("keydown", { key: "End", bubbles: true }));
|
||||
return {
|
||||
production,
|
||||
risky,
|
||||
keyboard: {
|
||||
selected: root.querySelector('[data-data-tab][aria-selected="true"]').dataset.dataTab,
|
||||
visible: root.querySelector("[data-data-view]:not([hidden])").dataset.dataView,
|
||||
},
|
||||
};
|
||||
})()`);
|
||||
|
||||
const layout = await evaluate(`(() => {
|
||||
const nav = document.querySelector(".top-nav");
|
||||
const meta = document.querySelector(".header-meta");
|
||||
return {
|
||||
documentOverflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
|
||||
navGap: Number((meta.getBoundingClientRect().left - nav.getBoundingClientRect().right).toFixed(1)),
|
||||
navLinks: document.querySelectorAll(".top-nav a").length,
|
||||
articleSections: document.querySelectorAll(".article-section").length,
|
||||
paperLinks: document.querySelectorAll(".paper-chain a").length,
|
||||
labTabs: document.querySelectorAll("[data-data-tab]").length,
|
||||
views: document.querySelectorAll("[data-data-view]").length,
|
||||
};
|
||||
})()`);
|
||||
|
||||
await evaluate(`(() => {
|
||||
document.documentElement.style.scrollBehavior = "auto";
|
||||
document.querySelector("[data-data-lab]").scrollIntoView({ block: "start", behavior: "instant" });
|
||||
})()`);
|
||||
await pause(150);
|
||||
await screenshot("/tmp/llm-atlas-data-lab-desktop.png");
|
||||
|
||||
await command("Emulation.setDeviceMetricsOverride", {
|
||||
width: 390,
|
||||
height: 844,
|
||||
deviceScaleFactor: 1,
|
||||
mobile: true,
|
||||
});
|
||||
await navigate("/pretraining/data/");
|
||||
await evaluate(`document.querySelector("[data-data-lab]").scrollIntoView({ block: "start", behavior: "instant" })`);
|
||||
await pause(120);
|
||||
await screenshot("/tmp/llm-atlas-data-lab-mobile.png");
|
||||
const mobile = await evaluate(`(() => {
|
||||
const toggle = document.querySelector("#menu-toggle");
|
||||
toggle.click();
|
||||
return {
|
||||
documentOverflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
|
||||
menuVisible: getComputedStyle(toggle).display !== "none",
|
||||
menuOpen: toggle.getAttribute("aria-expanded"),
|
||||
mobileLinks: document.querySelectorAll("#mobile-nav a").length,
|
||||
selectedTab: document.querySelector('[data-data-tab][aria-selected="true"]').dataset.dataTab,
|
||||
};
|
||||
})()`);
|
||||
|
||||
await command("Emulation.setDeviceMetricsOverride", {
|
||||
width: 1440,
|
||||
height: 1100,
|
||||
deviceScaleFactor: 1,
|
||||
mobile: false,
|
||||
});
|
||||
await navigate("/");
|
||||
const home = await evaluate(`({
|
||||
documentOverflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
|
||||
releaseCards: document.querySelectorAll(".release-card").length,
|
||||
firstRelease: document.querySelector(".release-card h2").textContent,
|
||||
firstHref: document.querySelector(".release-card").getAttribute("href"),
|
||||
paperCount: document.querySelector(".hero-stats div:nth-child(3) b").textContent,
|
||||
navLinks: document.querySelectorAll(".top-nav a").length,
|
||||
})`);
|
||||
|
||||
await navigate("/papers/");
|
||||
const papers = await evaluate(`(() => {
|
||||
const button = [...document.querySelectorAll("[data-filter]")].find((item) => item.dataset.filter === "数据");
|
||||
button?.click();
|
||||
return {
|
||||
hasDataFilter: Boolean(button),
|
||||
visible: [...document.querySelectorAll("[data-paper]")].filter((item) => !item.hidden).length,
|
||||
total: document.querySelectorAll("[data-paper]").length,
|
||||
};
|
||||
})()`);
|
||||
|
||||
const report = { pipeline, dedup, mixture, transform, layout, mobile, home, papers, exceptions };
|
||||
console.log(JSON.stringify(report, null, 2));
|
||||
|
||||
const numeric = (value) => Number.parseFloat(value);
|
||||
const failures = [];
|
||||
if (pipeline.balanced.rows !== 5 || pipeline.strict.rows !== 5) failures.push("Pipeline 漏斗阶段数量异常");
|
||||
if (numeric(pipeline.strict.retention) >= numeric(pipeline.balanced.retention)) failures.push("激进过滤没有降低文档保留率");
|
||||
if (dedup.exact.rows !== 6 || dedup.semantic.rows !== 6) failures.push("去重教学样本数量异常");
|
||||
if (numeric(dedup.semantic.recall) <= numeric(dedup.exact.recall) || Number(dedup.semantic.falsePositive) < 1) {
|
||||
failures.push("Semantic 去重没有体现召回—误伤权衡");
|
||||
}
|
||||
if (mixture.balanced.rows !== 5 || mixture.math.rows !== 5) failures.push("Mixture 五领域账本没有完整渲染");
|
||||
if (numeric(mixture.math.scores[2]) <= numeric(mixture.balanced.scores[2])) failures.push("数学强化没有提升教学 Math slice");
|
||||
if (numeric(transform.risky.errors) <= numeric(transform.production.errors) || transform.risky.errorDots <= transform.production.errorDots) {
|
||||
failures.push("失控合成没有增加错误副本");
|
||||
}
|
||||
if (transform.production.stages !== 4 || transform.risky.stages !== 4) failures.push("公开课程阶段数量异常");
|
||||
if (transform.keyboard.selected !== "transform" || transform.keyboard.visible !== "transform") failures.push("实验页签键盘导航异常");
|
||||
if (layout.articleSections !== 18 || layout.paperLinks !== 31 || layout.labTabs !== 4 || layout.views !== 4) {
|
||||
failures.push("章节、论文或实验数量异常");
|
||||
}
|
||||
if (layout.navLinks !== 13 || mobile.mobileLinks !== 13 || home.navLinks !== 13) failures.push("全站导航未同步数据工程");
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞:${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 6 || !home.firstRelease.includes("一万亿 Token") || home.firstHref !== "/pretraining/data/") {
|
||||
failures.push("首页数据工程新章入口异常");
|
||||
}
|
||||
if (home.paperCount !== "202") failures.push(`首页论文总数异常:${home.paperCount}`);
|
||||
if (!papers.hasDataFilter || papers.total !== 202 || papers.visible < 25) failures.push("论文库数据标签或论文总数异常");
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常:${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
if (failures.length) {
|
||||
failures.forEach((failure) => console.error(`- ${failure}`));
|
||||
process.exit(1);
|
||||
}
|
||||
console.log("数据工程专题真实 Chrome 断言通过。");
|
||||
@@ -177,7 +177,7 @@ if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentO
|
||||
}
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞:${layout.navGap}px`);
|
||||
if (!mobile.menuVisible) failures.push("移动端菜单按钮未显示");
|
||||
if (home.releaseCards !== 5) failures.push(`首页新章卡数量异常:${home.releaseCards}`);
|
||||
if (home.releaseCards !== 6) failures.push(`首页新章卡数量异常:${home.releaseCards}`);
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常:${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
@@ -289,7 +289,7 @@ if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentO
|
||||
}
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞:${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 5 || !home.firstRelease.includes("大一点")) failures.push("首页新章入口异常");
|
||||
if (home.releaseCards !== 6 || !home.firstRelease.includes("一万亿 Token")) failures.push("首页新章入口异常");
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常:${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
@@ -269,14 +269,14 @@ if (emergence.paths.some((length) => length < 500)) failures.push("涌现多指
|
||||
if (layout.articleSections !== 16 || layout.paperLinks !== 29 || layout.labTabs !== 4 || layout.views !== 4) {
|
||||
failures.push("章节、论文或实验数量异常");
|
||||
}
|
||||
if (layout.navLinks !== 12 || mobile.mobileLinks !== 12 || home.navLinks !== 12) failures.push("全站导航未同步 Scaling");
|
||||
if (layout.navLinks !== 13 || mobile.mobileLinks !== 13 || home.navLinks !== 13) failures.push("全站导航未同步 Scaling");
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞:${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 5 || !home.firstRelease.includes("大一点") || home.firstHref !== "/scaling/") {
|
||||
failures.push("首页 Scaling 新章入口异常");
|
||||
if (home.releaseCards !== 6 || !home.firstRelease.includes("一万亿 Token") || home.firstHref !== "/pretraining/data/") {
|
||||
failures.push("首页数据工程新章入口异常");
|
||||
}
|
||||
if (home.paperCount !== "182") failures.push(`首页论文总数异常:${home.paperCount}`);
|
||||
if (home.paperCount !== "202") failures.push(`首页论文总数异常:${home.paperCount}`);
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常:${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
@@ -233,7 +233,7 @@ if (layout.articleSections !== 16 || layout.paperLinks !== 37 || layout.labTabs
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞:${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 5 || !home.firstRelease.includes("大一点")) failures.push("首页新章入口异常");
|
||||
if (home.releaseCards !== 6 || !home.firstRelease.includes("一万亿 Token")) failures.push("首页新章入口异常");
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常:${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -11,6 +11,7 @@ const items = [
|
||||
{ id: "deepseek", href: "/deepseek/", label: "DeepSeek" },
|
||||
{ id: "foundations", href: "/foundations/", label: "基础原理" },
|
||||
{ id: "scaling", href: "/scaling/", label: "Scaling" },
|
||||
{ id: "data", href: "/pretraining/data/", label: "数据工程" },
|
||||
{ id: "moe", href: "/moe/", label: "MoE" },
|
||||
{ id: "long-context", href: "/long-context/", label: "长上下文" },
|
||||
{ id: "reasoning", href: "/reasoning/", label: "推理" },
|
||||
|
||||
@@ -86,12 +86,12 @@ export const chapters: Chapter[] = [
|
||||
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,
|
||||
summary: "用十二张账追踪采集、解析、过滤、去重、混合、改写、Tokenizer、Packing、课程与污染,并串起 DeepSeek / Kimi 数据谱系。",
|
||||
status: "published",
|
||||
progress: 73,
|
||||
papers: 31,
|
||||
prerequisites: ["01", "04"],
|
||||
highlights: ["数据质量", "混合策略", "污染"],
|
||||
highlights: ["十二张账", "DeepSeek/Kimi 谱系", "四联实验"],
|
||||
},
|
||||
{
|
||||
number: "06",
|
||||
|
||||
+182
-20
@@ -4,6 +4,7 @@ export type PaperTopic =
|
||||
| "长上下文"
|
||||
| "MoE"
|
||||
| "Scaling"
|
||||
| "数据"
|
||||
| "训练系统"
|
||||
| "低精度"
|
||||
| "后训练"
|
||||
@@ -123,8 +124,8 @@ export const papers: Paper[] = [
|
||||
year: 2019,
|
||||
title: "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer",
|
||||
url: "https://arxiv.org/abs/1910.10683",
|
||||
topics: ["Transformer", "Scaling"],
|
||||
contribution: "T5 统一 text-to-text 接口并系统研究架构与数据。",
|
||||
topics: ["Transformer", "Scaling", "数据"],
|
||||
contribution: "T5 统一 text-to-text 接口,并以 C4 建立现代 Common Crawl 清洗基线。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
@@ -397,7 +398,7 @@ export const papers: Paper[] = [
|
||||
year: 2021,
|
||||
title: "GLaM: Efficient Scaling of Language Models with Mixture-of-Experts",
|
||||
url: "https://arxiv.org/abs/2112.06905",
|
||||
topics: ["MoE", "Scaling"],
|
||||
topics: ["MoE", "Scaling", "数据"],
|
||||
contribution: "展示 MoE 在能耗与推理计算上高效扩展语言模型。",
|
||||
verified: true,
|
||||
},
|
||||
@@ -446,8 +447,8 @@ export const papers: Paper[] = [
|
||||
year: 2024,
|
||||
title: "DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model",
|
||||
url: "https://arxiv.org/abs/2405.04434",
|
||||
topics: ["MoE", "长上下文", "训练系统"],
|
||||
contribution: "联合 DeepSeekMoE 与 MLA,直接压缩推理 KV Cache。",
|
||||
topics: ["MoE", "长上下文", "数据", "训练系统"],
|
||||
contribution: "联合 DeepSeekMoE 与 MLA,并披露 8.1T 中英数据、过滤取舍和训练配方。",
|
||||
spotlight: "DeepSeek",
|
||||
verified: true,
|
||||
},
|
||||
@@ -472,7 +473,7 @@ export const papers: Paper[] = [
|
||||
year: 2024,
|
||||
title: "DeepSeek-V3 Technical Report",
|
||||
url: "https://arxiv.org/abs/2412.19437",
|
||||
topics: ["MoE", "Scaling", "训练系统", "低精度"],
|
||||
topics: ["MoE", "Scaling", "数据", "训练系统", "低精度"],
|
||||
contribution: "671B-A37B;FP8、MTP、无辅助损失平衡与 DualPipe。",
|
||||
spotlight: "DeepSeek",
|
||||
verified: true,
|
||||
@@ -545,8 +546,8 @@ export const papers: Paper[] = [
|
||||
year: 2024,
|
||||
title: "DeepSeek LLM: Scaling Open-Source Language Models with Longtermism",
|
||||
url: "https://arxiv.org/abs/2401.02954",
|
||||
topics: ["Scaling"],
|
||||
contribution: "中英 dense 基线与公开 scaling-law 研究。",
|
||||
topics: ["Scaling", "数据"],
|
||||
contribution: "中英 dense 基线、跨 91 个 Common Crawl dump 去重与公开 scaling-law 研究。",
|
||||
spotlight: "DeepSeek",
|
||||
verified: true,
|
||||
},
|
||||
@@ -554,7 +555,7 @@ export const papers: Paper[] = [
|
||||
year: 2021,
|
||||
title: "The Pile: An 800GB Dataset of Diverse Text for Language Modeling",
|
||||
url: "https://arxiv.org/abs/2101.00027",
|
||||
topics: ["Scaling"],
|
||||
topics: ["Scaling", "数据"],
|
||||
contribution: "公开多域预训练语料与数据配比研究基础。",
|
||||
verified: true,
|
||||
},
|
||||
@@ -562,7 +563,7 @@ export const papers: Paper[] = [
|
||||
year: 2023,
|
||||
title: "The RefinedWeb Dataset for Falcon LLM",
|
||||
url: "https://arxiv.org/abs/2306.01116",
|
||||
topics: ["Scaling"],
|
||||
topics: ["Scaling", "数据"],
|
||||
contribution: "证明高质量过滤与去重的 Web 数据可以支撑强模型。",
|
||||
verified: true,
|
||||
},
|
||||
@@ -570,10 +571,170 @@ export const papers: Paper[] = [
|
||||
year: 2024,
|
||||
title: "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research",
|
||||
url: "https://arxiv.org/abs/2402.00159",
|
||||
topics: ["Scaling"],
|
||||
topics: ["Scaling", "数据"],
|
||||
contribution: "开放 3T Token 语料与完整数据处理工具。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2019,
|
||||
title: "CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data",
|
||||
url: "https://arxiv.org/abs/1911.00359",
|
||||
topics: ["数据"],
|
||||
contribution: "把去重、语言识别和 Wikipedia-like 质量过滤组织成可扩展网页语料流水线。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2020,
|
||||
title: "Extracting Training Data from Large Language Models",
|
||||
url: "https://arxiv.org/abs/2012.07805",
|
||||
topics: ["数据", "评测"],
|
||||
contribution: "从 GPT-2 抽取逐字训练片段与 PII,实证训练记忆和隐私泄露风险。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2021,
|
||||
title: "Deduplicating Training Data Makes Language Models Better",
|
||||
url: "https://arxiv.org/abs/2107.06499",
|
||||
topics: ["数据", "评测"],
|
||||
contribution: "连接近重复、验证集重合、训练效率与逐字记忆,建立现代 fuzzy dedup 主线。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2023,
|
||||
title: "ROOTS: One Massive Multilingual Dataset for 1.6TB of Text",
|
||||
url: "https://arxiv.org/abs/2303.03915",
|
||||
topics: ["数据"],
|
||||
contribution: "总结 59 种语言、1.6TB 多语言语料的协作建设、治理与数据文档。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2022,
|
||||
title: "Efficient Training of Language Models to Fill in the Middle",
|
||||
url: "https://arxiv.org/abs/2207.14255",
|
||||
topics: ["数据"],
|
||||
contribution: "通过 PSM/SPM 数据重排获得代码中间补全能力,并消融变换率和格式。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2023,
|
||||
title: "DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining",
|
||||
url: "https://arxiv.org/abs/2305.10429",
|
||||
topics: ["数据", "Scaling"],
|
||||
contribution: "用小代理模型学习 domain weights,再迁移到更大目标模型。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2023,
|
||||
title: "SemDeDup: Data-efficient Learning at Web-scale through Semantic Deduplication",
|
||||
url: "https://arxiv.org/abs/2303.09540",
|
||||
topics: ["数据"],
|
||||
contribution: "在预训练 embedding 空间聚类语义近重复,研究硬删除与多样性收益。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2023,
|
||||
title: "D4: Improving LLM Pretraining via Document De-Duplication and Diversification",
|
||||
url: "https://arxiv.org/abs/2308.12284",
|
||||
topics: ["数据"],
|
||||
contribution: "先去重再主动保留多样性,说明智能重复与随机重复并不等价。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2023,
|
||||
title: "PALOMA: A Benchmark for Evaluating Language Model Fit",
|
||||
url: "https://arxiv.org/abs/2312.10523",
|
||||
topics: ["数据", "评测"],
|
||||
contribution: "提出固定训练顺序、子文档去污染、细粒度 domain 和 bits-per-byte 语料评测协议。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale",
|
||||
url: "https://arxiv.org/abs/2406.17557",
|
||||
topics: ["数据"],
|
||||
contribution: "以大规模网页清洗消融构建 FineWeb 与教育价值子集。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "DataComp-LM: In Search of the Next Generation of Language Model Pretraining Datasets",
|
||||
url: "https://arxiv.org/abs/2406.11794",
|
||||
topics: ["数据", "Scaling", "评测"],
|
||||
contribution: "固定 240T candidate pool、模型尺度与 53 项评测,使数据 pipeline 可以同协议比较。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling",
|
||||
url: "https://arxiv.org/abs/2401.16380",
|
||||
topics: ["数据"],
|
||||
contribution: "WRAP 用受控改写增加网页语料的表达覆盖,并研究 Token utility。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "Fewer Truncations Improve Language Modeling",
|
||||
url: "https://arxiv.org/abs/2404.10830",
|
||||
topics: ["数据", "训练系统"],
|
||||
contribution: "Best-fit packing 减少长文档截断,在固定序列长度下保护文档完整性。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "SoftDedup: Efficient Data Reweighting at Scale",
|
||||
url: "https://arxiv.org/abs/2407.06654",
|
||||
topics: ["数据"],
|
||||
contribution: "按常见 n-gram 频率降低样本权重,探索比硬删除更连续的去重策略。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "MATES: Model-Aware Data Selection for Efficient Language Model Pretraining",
|
||||
url: "https://arxiv.org/abs/2406.06046",
|
||||
topics: ["数据", "Scaling"],
|
||||
contribution: "以当前模型状态动态估计样本对目标验证分布的影响。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "Investigating Data Contamination in Modern Benchmarks for Large Language Models",
|
||||
url: "https://arxiv.org/abs/2401.06059",
|
||||
topics: ["数据", "评测"],
|
||||
contribution: "区分 text contamination 与 ground-truth contamination,并检查 n-gram 检测的边界。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training",
|
||||
url: "https://arxiv.org/abs/2403.09611",
|
||||
topics: ["数据", "多模态"],
|
||||
contribution: "系统消融 caption、interleaved image-text、text-only mixture 与视觉组件。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2024,
|
||||
title: "JEST: Joint Example Selection for Multimodal Learning",
|
||||
url: "https://arxiv.org/abs/2406.17711",
|
||||
topics: ["数据", "多模态"],
|
||||
contribution: "按 batch 的联合可学习性选择多模态样本,研究动态数据价值。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2025,
|
||||
title: "Group-MATES: Data-Efficient Pretraining with Group-Level Data Influence Modeling",
|
||||
url: "https://arxiv.org/abs/2502.14709",
|
||||
topics: ["数据", "Scaling"],
|
||||
contribution: "把模型感知的数据影响估计扩展到 group-level 选择。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2025,
|
||||
title: "Swallow: Continued Pre-training with Synthetic Data for Efficient Language Model Adaptation",
|
||||
url: "https://arxiv.org/abs/2505.02881",
|
||||
topics: ["数据", "后训练"],
|
||||
contribution: "用 SwallowCode 与 SwallowMath 研究受控合成数据的继续预训练收益与边界。",
|
||||
verified: true,
|
||||
},
|
||||
{
|
||||
year: 2017,
|
||||
title: "Deep Learning Scaling is Predictable, Empirically",
|
||||
@@ -1157,8 +1318,8 @@ export const papers: Paper[] = [
|
||||
year: 2024,
|
||||
title: "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models",
|
||||
url: "https://arxiv.org/abs/2402.03300",
|
||||
topics: ["推理", "后训练"],
|
||||
contribution: "数学数据管线与无独立 critic 的 GRPO。",
|
||||
topics: ["数据", "推理", "后训练"],
|
||||
contribution: "从 40B HTML pages 发现 120.2B 数学语料,并提出无独立 critic 的 GRPO。",
|
||||
spotlight: "DeepSeek",
|
||||
verified: true,
|
||||
},
|
||||
@@ -1344,8 +1505,8 @@ export const papers: Paper[] = [
|
||||
year: 2025,
|
||||
title: "Kimi K2: Open Agentic Intelligence",
|
||||
url: "https://arxiv.org/abs/2507.20534",
|
||||
topics: ["MoE", "训练系统", "Agent", "后训练", "推理"],
|
||||
contribution: "开放 1T MoE Agent 模型,以 verifiable gym 与 self-critique rubric 做 joint RL。",
|
||||
topics: ["MoE", "数据", "训练系统", "Agent", "后训练", "推理"],
|
||||
contribution: "开放 1T MoE Agent 模型,以 15.5T、知识/数学改写与 verifiable gym / self-critique joint RL 协同。",
|
||||
spotlight: "Kimi",
|
||||
verified: true,
|
||||
},
|
||||
@@ -1353,7 +1514,7 @@ export const papers: Paper[] = [
|
||||
year: 2026,
|
||||
title: "Kimi K2.5: Visual Agentic Intelligence",
|
||||
url: "https://arxiv.org/abs/2602.02276",
|
||||
topics: ["训练系统", "Agent", "多模态", "推理"],
|
||||
topics: ["数据", "训练系统", "Agent", "多模态", "推理"],
|
||||
contribution: "把视觉、工具使用和并行 Agent Swarm 纳入统一后训练。",
|
||||
spotlight: "Kimi",
|
||||
verified: true,
|
||||
@@ -1484,8 +1645,8 @@ export const papers: Paper[] = [
|
||||
year: 2026,
|
||||
title: "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence",
|
||||
url: "https://arxiv.org/abs/2606.19348",
|
||||
topics: ["长上下文", "MoE", "训练系统", "推理"],
|
||||
contribution: "CSA/HCA 混合注意力、mHC、Muon 与 1M 上下文双模型。",
|
||||
topics: ["长上下文", "MoE", "数据", "训练系统", "推理"],
|
||||
contribution: "32/33T 长文与 agentic 数据、CSA/HCA、mHC、Muon 和 1M 上下文双模型。",
|
||||
spotlight: "DeepSeek",
|
||||
verified: true,
|
||||
},
|
||||
@@ -1493,8 +1654,8 @@ export const papers: Paper[] = [
|
||||
year: 2026,
|
||||
title: "Kimi K3: Open Frontier Intelligence",
|
||||
url: "https://arxiv.org/abs/2607.24653",
|
||||
topics: ["长上下文", "MoE", "训练系统", "低精度", "推理", "Agent", "多模态", "评测"],
|
||||
contribution: "2.8T-A104B、KDA/MLA、AttnRes、Stable LatentMoE 与 1M Agentic RL。",
|
||||
topics: ["长上下文", "MoE", "数据", "训练系统", "低精度", "推理", "Agent", "多模态", "评测"],
|
||||
contribution: "四文本域与视觉语料、2.8T-A104B、KDA/MLA、Stable LatentMoE 与 1M Agentic RL。",
|
||||
spotlight: "Kimi",
|
||||
verified: true,
|
||||
},
|
||||
@@ -1506,6 +1667,7 @@ export const paperTopics: PaperTopic[] = [
|
||||
"长上下文",
|
||||
"MoE",
|
||||
"Scaling",
|
||||
"数据",
|
||||
"训练系统",
|
||||
"低精度",
|
||||
"后训练",
|
||||
|
||||
+29
-2
@@ -8,6 +8,7 @@ const routes: Record<string, string> = {
|
||||
roadmap: "/roadmap/",
|
||||
foundations: "/foundations/",
|
||||
scaling: "/scaling/",
|
||||
"pretraining/data": "/pretraining/data/",
|
||||
moe: "/moe/",
|
||||
"long-context": "/long-context/",
|
||||
reasoning: "/reasoning/",
|
||||
@@ -80,6 +81,7 @@ const paths = [
|
||||
<a class="button" href="/k3/">直接解剖 K3</a>
|
||||
<a class="button" href="/deepseek/">DeepSeek 专题</a>
|
||||
<a class="button" href="/scaling/">Scaling Laws 专题</a>
|
||||
<a class="button" href="/pretraining/data/">数据工程专题</a>
|
||||
<a class="button" href="/moe/">MoE 专题</a>
|
||||
<a class="button" href="/long-context/">长上下文专题</a>
|
||||
<a class="button" href="/reasoning/">推理专题</a>
|
||||
@@ -93,20 +95,36 @@ const paths = [
|
||||
<div class="hero-stats">
|
||||
<div><b>16</b><span>核心专题</span></div>
|
||||
<div><b>151</b><span>K3 报告来源</span></div>
|
||||
<div><b>182</b><span>关键论文索引</span></div>
|
||||
<div><b>202</b><span>关键论文索引</span></div>
|
||||
<div><b>47p</b><span>K3 技术报告</span></div>
|
||||
</div>
|
||||
</aside>
|
||||
</div>
|
||||
<div class="scope-strip">
|
||||
<span>研究范围</span>
|
||||
<p>表示学习 · Transformer · Scaling · MoE · 长上下文 · 训练系统 · 后训练 · 推理 · Agent · 多模态 · 服务与评测</p>
|
||||
<p>表示学习 · Transformer · Scaling · 数据工程 · MoE · 长上下文 · 训练系统 · 后训练 · 推理 · Agent · 多模态 · 服务与评测</p>
|
||||
<a href="/progress/">查看实时进度 →</a>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="section compact release-section" id="new-chapters">
|
||||
<div class="release-grid">
|
||||
<a class="release-card data-release" href="/pretraining/data/">
|
||||
<div>
|
||||
<p class="eyebrow"><span>NEW / CHAPTER 05</span> DATA & PRE-TRAINING</p>
|
||||
<h2>同样一万亿 Token,可以是完全不同的训练经历</h2>
|
||||
<p>
|
||||
用来源、解析、质量、唯一性、污染、混合、变换、Tokenizer、Packing、课程与模型感知价值十二张账,
|
||||
从 C4、Pile、DoReMi 和 DCLM 一路走到 DeepSeek 五代数据工程与 Kimi K2→K3 原生多模态。
|
||||
</p>
|
||||
</div>
|
||||
<dl>
|
||||
<div><dt>LINEAGE</dt><dd>2019 → 2026</dd></div>
|
||||
<div><dt>PAPERS</dt><dd>31 个一手节点</dd></div>
|
||||
<div><dt>LAB</dt><dd>流水线 · 去重 · 混合 · 改写</dd></div>
|
||||
</dl>
|
||||
<span class="release-arrow" aria-hidden="true">进入数据工程专题 →</span>
|
||||
</a>
|
||||
<a class="release-card scaling-release" href="/scaling/">
|
||||
<div>
|
||||
<p class="eyebrow"><span>NEW / CHAPTER 04</span> SCALING LAWS</p>
|
||||
@@ -420,6 +438,7 @@ const paths = [
|
||||
transition: transform 180ms ease, border-color 180ms ease;
|
||||
}
|
||||
|
||||
.data-release,
|
||||
.scaling-release,
|
||||
.training-release,
|
||||
.reasoning-release {
|
||||
@@ -427,6 +446,13 @@ const paths = [
|
||||
min-height: 510px;
|
||||
}
|
||||
|
||||
.data-release {
|
||||
background:
|
||||
radial-gradient(circle at 82% 18%, rgba(76, 118, 112, 0.2), transparent 30%),
|
||||
repeating-linear-gradient(0deg, transparent 0 46px, rgba(76, 118, 112, 0.045) 46px 47px),
|
||||
var(--paper-raised);
|
||||
}
|
||||
|
||||
.scaling-release {
|
||||
background:
|
||||
radial-gradient(circle at 82% 18%, rgba(159, 91, 52, 0.18), transparent 30%),
|
||||
@@ -510,6 +536,7 @@ const paths = [
|
||||
padding-bottom: 76px;
|
||||
}
|
||||
|
||||
.data-release,
|
||||
.scaling-release,
|
||||
.training-release,
|
||||
.reasoning-release {
|
||||
|
||||
@@ -683,7 +683,7 @@ const paperChain = [
|
||||
<div class="next-links">
|
||||
<a class="button primary" href="/k3/#moe">回到 K3:在整机架构中定位 Stable LatentMoE →</a>
|
||||
<a class="button" href="/deepseek/">DeepSeek 完整论文谱系</a>
|
||||
<a class="button" href="/papers/">搜索全部 182 篇论文</a>
|
||||
<a class="button" href="/papers/">搜索全部 202 篇论文</a>
|
||||
</div>
|
||||
</section>
|
||||
</article>
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -12,6 +12,7 @@ const workstreams = [
|
||||
{ label: "Kimi K3 深读", value: 66, next: "扩写 pre-training / infra 逐图笔记" },
|
||||
{ label: "Transformer 基础", value: 52, next: "加入矩阵形状动画与手算练习" },
|
||||
{ label: "Scaling Laws", value: 74, next: "加入真实拟合复现、置信区间与更多模型族对照" },
|
||||
{ label: "数据工程与预训练配方", value: 73, next: "逐图精读 FineWeb / DCLM,加入真实去重与 mixture traces" },
|
||||
{ label: "DeepSeek 专题", value: 71, next: "补 R1 / DAPO 的逐图训练轨迹与复现对照" },
|
||||
{ label: "推理与测试时扩展", value: 76, next: "真实模型采样曲线、PRM 案例与逐篇图表精读" },
|
||||
{ label: "稀疏计算与 MoE", value: 74, next: "补充真实集群 traces 与专家特化案例" },
|
||||
@@ -41,7 +42,7 @@ const workstreams = [
|
||||
<div><dt>OVERALL</dt><dd>专题平均 {average}%</dd></div>
|
||||
<div><dt>READABLE</dt><dd>{published} 个首版可读专题</dd></div>
|
||||
<div><dt>ACTIVE</dt><dd>{researching} 个研究/写作中</dd></div>
|
||||
<div><dt>UPDATED</dt><dd>2026-07-29 02:08 CST</dd></div>
|
||||
<div><dt>UPDATED</dt><dd>2026-07-29 03:01 CST</dd></div>
|
||||
<div><dt>MODE</dt><dd>持续迭代,不锁死版本</dd></div>
|
||||
</dl>
|
||||
</div>
|
||||
@@ -51,7 +52,7 @@ const workstreams = [
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>01</span> WORKSTREAMS</p>
|
||||
<h2>十二条工作流同时推进,但不混淆“有页面”和“已核验”</h2>
|
||||
<h2>十三条工作流同时推进,但不混淆“有页面”和“已核验”</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
内容首版优先打通全局脉络;随后每轮迭代选择一个专题推进到论文/工程层,并做独立事实复核。
|
||||
@@ -88,14 +89,15 @@ const workstreams = [
|
||||
<article><span>✓</span><h3>K3 报告已结构化拆解</h3><p>47 页报告目录、151 条参考来源和架构/后训练/系统主线已经提取。</p></article>
|
||||
<article><span>✓</span><h3>16 专题知识图</h3><p>从语言模型基础到评测安全,包含先修依赖和三条贯穿案例。</p></article>
|
||||
<article><span>✓</span><h3>编辑式网站系统</h3><p>响应式导航、章节模板、侧栏、进度、论文链和证据提示组件。</p></article>
|
||||
<article><span>✓</span><h3>十六个原创交互视图</h3><p>K3、注意力、DeepSeek、长上下文、MoE、推理三页签,以及训练系统与 Scaling 各四页签实验。</p></article>
|
||||
<article><span>✓</span><h3>八篇首版长文</h3><p>K3 导读、Transformer 基础、DeepSeek 谱系、Scaling、长上下文、MoE、推理与训练系统专题。</p></article>
|
||||
<article><span>✓</span><h3>二十个原创交互视图</h3><p>K3、注意力、DeepSeek、长上下文、MoE、推理,以及训练系统、Scaling 与数据工程的多页签实验。</p></article>
|
||||
<article><span>✓</span><h3>九篇首版长文</h3><p>K3、Transformer、DeepSeek、Scaling、数据工程、长上下文、MoE、推理与训练系统专题。</p></article>
|
||||
<article><span>✓</span><h3>Scaling Laws 深度专题</h3><p>九张账、29 个一手节点、DeepSeek/Kimi 双谱系与曲面—部署—复用—涌现四联实验。</p></article>
|
||||
<article><span>✓</span><h3>数据工程深度专题</h3><p>十二张账、31 个一手节点、DeepSeek/Kimi 双谱系与流水线—去重—混合—改写四联实验。</p></article>
|
||||
<article><span>✓</span><h3>长上下文深度专题</h3><p>五张成本账、26 篇一手论文、10+ 机制图与 8 策略交互实验室。</p></article>
|
||||
<article><span>✓</span><h3>MoE 深度专题</h3><p>六张账、19 篇一手论文、DeepSeek/K3 主线与路由—容量—通信交互实验室。</p></article>
|
||||
<article><span>✓</span><h3>推理深度专题</h3><p>八张账、30 篇一手论文链、DeepSeek/Kimi 双主线与三页签互动实验室。</p></article>
|
||||
<article><span>✓</span><h3>训练系统深度专题</h3><p>九张账、37 个一手节点、DeepSeek/Kimi 双谱系与显存—网格—气泡—通信实验室。</p></article>
|
||||
<article><span>✓</span><h3>182 篇关键论文索引</h3><p>覆盖 12 个专题,支持全文搜索、标签筛选与 Kimi/DeepSeek 聚光主线。</p></article>
|
||||
<article><span>✓</span><h3>202 篇关键论文索引</h3><p>新增数据专题标签和 20 个一手节点,继续支持全文搜索与 Kimi/DeepSeek 聚光主线。</p></article>
|
||||
<article><span>✓</span><h3>公开仓库与自托管发布</h3><p>源码公开到 git.k1412.top,网站由不可变镜像、Compose Manager 与 HTTPS 交付。</p></article>
|
||||
</div>
|
||||
</section>
|
||||
@@ -111,6 +113,7 @@ const workstreams = [
|
||||
<div class="queue-table">
|
||||
<div class="head"><b>优先级</b><b>专题</b><b>本轮交付</b><b>完成闸门</b></div>
|
||||
<div><span>P0</span><strong>Scaling Laws 二轮</strong><p>真实拟合复现 → 置信区间 → 更多模型族与下游任务外推</p><em>可复现实验 + 逐图笔记</em></div>
|
||||
<div><span>P1</span><strong>数据工程二轮</strong><p>FineWeb / DCLM 逐图 → 真实去重误伤 → mixture traces 与污染案例</p><em>逐图笔记 + 案例库</em></div>
|
||||
<div><span>P1</span><strong>大规模训练系统二轮</strong><p>真实集群 traces → 故障恢复 → 精确 topology / kernel 配置</p><em>案例库 + 实测边界</em></div>
|
||||
<div><span>P1</span><strong>长上下文二轮深化</strong><p>真实模型配置 → 内核细节 → 长上下文评测与失败案例</p><em>配置比较器 + 逐图论文笔记</em></div>
|
||||
<div><span>P1</span><strong>MoE 二轮深化</strong><p>真实负载 traces → 专家特化可解释性 → 共享专家语义</p><em>案例库 + 集群证据</em></div>
|
||||
@@ -157,6 +160,8 @@ const workstreams = [
|
||||
<div><time>2026-07-29</time><b>推理按八张账组织</b><p>把答案、覆盖、选择、过程、预算、优化、分布与系统证据分开核算。</p></div>
|
||||
<div><time>2026-07-29</time><b>训练系统按九张账组织</b><p>把模型状态、激活、并行、气泡、通信、专家、上下文、数值与可靠性分开核算。</p></div>
|
||||
<div><time>2026-07-29</time><b>Scaling Laws 按九张账组织</b><p>把观测量、模型、数据、算术、配比、配方、外推、经济目标与能力阶段分开核算。</p></div>
|
||||
<div><time>2026-07-29</time><b>数据工程按十二张账组织</b><p>把来源、解析、语言、质量、唯一性、污染、混合、变换、Tokenizer、Packing、课程与价值分开核算。</p></div>
|
||||
<div><time>2026-07-29</time><b>Grok 数据线索与正式账本永久分离</b><p>1264 行正式研究账本只接受回查一手来源后的结论;203 行 Grok 产物保留为未核验发现队列。</p></div>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
|
||||
@@ -699,7 +699,7 @@ const paperChain = [
|
||||
<h2>不要按模型榜单读,按“旧方法留下了哪堵墙”读</h2>
|
||||
<p>
|
||||
下面 30 篇构成本专题首版主干。课程中的机制、公式与数字优先回到这些一手来源;
|
||||
更完整的 182 篇跨专题索引位于论文库。
|
||||
更完整的 202 篇跨专题索引位于论文库。
|
||||
</p>
|
||||
<div class="paper-chain">
|
||||
{paperChain.map(([year, title, href, note], index) => (
|
||||
|
||||
Reference in New Issue
Block a user