Skip to content

Commit 0ee40fa

Browse files
longsizhuolongsizhuo
committed
feat(docs): i18n ai/ 后半部分翻译完成 + MDX 语法修复 (32 篇)
translator-ai-2 产出 + MDX 语法修复。全部 zh→en。 覆盖： - llm-basics (12 篇): courses / cuda / deep-learning / embeddings / pytorch / transformer - methodology (1) - misc-tools (1) - model-datasets-platforms (1) - multimodal (10): courses / llava / mllm / qwenvl / ViT / VAE / VQVAE / RQVAE 等 - recommender-systems (7): 王树森推荐系统笔记全集（粗排/精排/重排/冷启动等） MDX 语法修复： - 2 个 recommender-systems 文件的裸 <1000 / <30 / <24 / <8 用反引号包裹 （MDX 把 <digit 当成 JSX tag 开头导致 build 失败） - app/api/analytics/top-docs/route.ts: Prisma JSON filter startsWith 写法错误，改为内存筛选 术语决策（新发现）: - 笔记 → post (Xiaohongshu 场景) - 粗排/精排/重排 → pre-ranking / full ranking / re-ranking - 保量 → exposure guarantee - 融合分数 → fused score - 老汤模型 → aged model (inline explanation)
1 parent 35c96b5 commit 0ee40fa

33 files changed

Lines changed: 6007 additions & 3 deletions

‎app/api/analytics/top-docs/route.ts‎

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -17,21 +17,24 @@ export async function GET(req: NextRequest) {
1717
since.setFullYear(since.getFullYear() - 10);
1818
}
1919

20+
// Prisma 对 JSON 字段的 startsWith 过滤不能直接嵌套写在 where，
21+
// 这里先按 eventType + createdAt 过滤，再在内存里按 path 前缀筛
2022
const rows = await prisma.analyticsEvent.findMany({
2123
where: {
2224
eventType: "page_view",
2325
createdAt: { gte: since },
24-
eventData: { path: { startsWith: "/docs/" } },
2526
},
2627
select: { eventData: true },
2728
});
2829

29-
// 统计各路径 PV
30+
// 统计各路径 PV（内存过滤 /docs/ 前缀）
3031
const counts: Record<string, number> = {};
3132
for (const row of rows) {
3233
const data = row.eventData as { path?: string; title?: string } | null;
3334
const path = data?.path;
34-
if (path) counts[path] = (counts[path] ?? 0) + 1;
35+
if (path && path.startsWith("/docs/")) {
36+
counts[path] = (counts[path] ?? 0) + 1;
37+
}
3538
}
3639

3740
const top = Object.entries(counts)

‎app/docs/ai/llm-basics/courses/index.en.mdx‎

Lines changed: 70 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,70 @@
1+
---
2+
title: LLM Introductory Courses
3+
description: A curated collection of courses on deep learning and large language models.
4+
docId: xboc8qj2128aivvt0goo1wow
5+
lang: en
6+
translatedFrom: zh
7+
translatedAt: 2026-04-15T12:00:00Z
8+
translatorAgent: claude-sonnet-4-6
9+
---
10+
11+
# Introductory Courses
12+
13+
## CS224N — Stanford Deep Learning for NLP
14+
15+
Covers NLP fundamentals and includes a miniGPT project.
16+
17+
- Official site: [CS224N Stanford](https://web.stanford.edu/class/cs224n/index.html#schedule)
18+
- Video: (2025, bilingual) Stanford CS224N — Deep Learning for Natural Language Processing
19+
- Slides and assignments: [Quark Cloud Drive](https://pan.quark.cn/s/1e43e9b25006)
20+
21+
---
22+
23+
## CMU Advanced NLP
24+
25+
- Course homepage: [CMU Advanced NLP Spring 2025](https://cmu-l3.github.io/anlp-spring2025/)
26+
- Code: [GitHub Repo](https://github.com/cmu-l3/anlp-spring2025-code/tree/main)
27+
- Assignments:
28+
- **minLLaMA**: [HW1](https://github.com/cmu-l3/anlp-spring2025-hw1) (implement LLaMA from scratch — extremely high quality)
29+
- **RAG**: [HW2](https://github.com/cmu-l3/anlp-spring2025-hw2)
30+
31+
---
32+
33+
## NanoGPT — GPT from Scratch
34+
35+
- Code: [NanoGPT GitHub](https://github.com/karpathy/nanoGPT)
36+
- Learning materials: [LLM Training Series] NanoGPT source code walkthrough and Chinese GPT training practice
37+
38+
---
39+
40+
## Stanford CS336 — Language Modeling from Scratch (Spring 2025)
41+
42+
- Course videos: [YouTube Playlist](https://www.youtube.com/watch?v=SQ3fZ1sAqXI&list=PLoROMvodv4rOY23Y0BoGoBGgQ1zmU_MT_)
43+
- Course homepage: [CS336 Official Site](https://stanford-cs336.github.io/spring2025/)
44+
- Code repository: [GitHub Repo](https://github.com/stanford-cs336/spring2025-lectures)
45+
- Video translation: [Bilibili](https://www.bilibili.com/video/BV13SV9zdEhX/?spm_id_from=333.337.search-card.all.click&vd_source=14245d272f6606a31fe299db9e47ca84)
46+
- Translation collection: [Zhihu Column](https://zhuanlan.zhihu.com/p/1906315844386034284)
47+
- Private assignment repo: TODO
48+
49+
### Assignment Overview
50+
51+
1. **Assignment 1**: Implement a BPE tokenizer, Transformer architecture, and Adam optimizer; train on TinyStories and OpenWebText (PyTorch primitives only).
52+
2. **Assignment 2**: Implement Flash Attention 2 in Triton; distributed data parallelism + optimizer sharding.
53+
3. **Assignment 3**: Scaling Laws. Fit scaling laws using IsoFLOP, simulating experiments under a fixed compute budget.
54+
4. **Assignment 4**: Data pipeline. Convert Common Crawl HTML to text, filter (quality, harmful content, PII), deduplicate.
55+
5. **Assignment 5**: Alignment. Implement supervised fine-tuning, expert iteration, GRPO and variants; run RL on Qwen 2.5 Math 1.5B to improve MATH benchmark performance.
56+
57+
### Prerequisites
58+
59+
- Math: MATH 51, CME 100
60+
- Probability: CS 109
61+
62+
### Study Notes
63+
64+
- CS336 study notes: TODO
65+
66+
---
67+
68+
## Happy-LLM — Build a 215M LLM from Scratch
69+
70+
- Code repository: [Happy-LLM GitHub](https://github.com/datawhalechina/happy-llm)

‎app/docs/ai/llm-basics/cuda/index.en.mdx‎

Lines changed: 34 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,34 @@
1+
---
2+
title: CUDA
3+
description: CUDA learning resources and large model optimization techniques
4+
date: "2024-01-17"
5+
tags:
6+
- cuda
7+
- gpu
8+
- triton
9+
- cutlass
10+
- flashattention
11+
- ring-attention
12+
- profiling
13+
docId: nwt5322vw4q6sz8ho8qynv28
14+
lang: en
15+
translatedFrom: zh
16+
translatedAt: 2026-04-15T12:00:00Z
17+
translatorAgent: claude-sonnet-4-6
18+
---
19+
20+
## Recommended Courses
21+
22+
- [CUDA Lectures (comprehensive GitHub course)](https://github.com/cuda-mode/lectures)
23+
Covers **profiling, Triton, Cutlass, FlashAttention, Ring Attention** and other practical topics.
24+
25+
## Related Articles
26+
27+
- [Zhihu Column: Advanced CUDA Learning](https://zhuanlan.zhihu.com/p/711304830)
28+
29+
## Learning Value
30+
31+
- Understand GPU parallel computing and memory optimization
32+
- Master the core components of model acceleration (FlashAttention, Ring Attention)
33+
- Learn to use profiling tools to identify performance bottlenecks
34+
- Become familiar with next-generation high-performance GPU programming frameworks such as Triton and Cutlass

‎app/docs/ai/llm-basics/deep-learning/d2l/index.en.mdx‎

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
---
2+
title: Dive into Deep Learning (D2L) by Mu Li
3+
description: Notes related to Dive into Deep Learning.
4+
date: "2024-01-18"
5+
tags:
6+
- d2l
7+
- deep-learning
8+
- pytorch
9+
- notes
10+
docId: dqg4iqz7hgyq38cqz3tg9tlf
11+
lang: en
12+
translatedFrom: zh
13+
translatedAt: 2026-04-15T12:00:00Z
14+
translatorAgent: claude-sonnet-4-6
15+
---
16+
17+
## Official Resources
18+
19+
- [Dive into Deep Learning (D2L Official Site)](https://zh-v2.d2l.ai/)
20+
21+
## E-Books
22+
23+
- [Mu Li — Dive into Deep Learning.pdf]
24+
- [Dive into Deep Learning (PyTorch Edition).pdf]
25+
26+
## Study Notes
27+
28+
- Dive into Deep Learning notes — TODO
29+
- Quark Cloud Drive: [D2L Chinese Notes](https://pan.quark.cn/s/9a7cf3f3eae2)

‎app/docs/ai/llm-basics/deep-learning/index.en.mdx‎

Lines changed: 111 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,111 @@
1+
---
2+
title: Deep Learning Fundamentals
3+
description: "Deep learning fundamentals: Dive into Deep Learning by Mu Li, NLP, and machine learning resources"
4+
date: "2025-01-27"
5+
tags:
6+
- deep-learning
7+
- d2l
8+
- nlp
9+
- machine-learning
10+
docId: vdclex41huib10ccsqw9u76k
11+
lang: en
12+
translatedFrom: zh
13+
translatedAt: 2026-04-15T12:00:00Z
14+
translatorAgent: claude-sonnet-4-6
15+
---
16+
17+
Deep learning is the theoretical foundation of large language models. This section provides systematic learning resources and practical guidance.
18+
19+
## Dive into Deep Learning by Mu Li
20+
21+
### Core Resources
22+
23+
- **Official website**: [https://zh-v2.d2l.ai/](https://zh-v2.d2l.ai/) — Chinese online tutorial
24+
- **Highlights**: Equal emphasis on theory and code; provides both PyTorch and MXNet implementations
25+
- **Coverage**: From basic linear regression to advanced attention mechanisms
26+
27+
### Learning Materials
28+
29+
- **PDF edition**: Mu Li — Dive into Deep Learning
30+
- **PyTorch edition**: Dive into Deep Learning (PyTorch Edition)
31+
- **Notes**: Dive into Deep Learning Chinese Notes
32+
- Quark Cloud Drive: https://pan.quark.cn/s/9a7cf3f3eae2
33+
34+
### Characteristics
35+
36+
- **Practice-oriented**: Every concept has a corresponding code implementation
37+
- **Progressive**: Builds from simple concepts to complex models step by step
38+
- **Comprehensive**: Covers the main areas of deep learning
39+
- **Up-to-date**: Continuously updated with the latest techniques and methods
40+
41+
## Learning Recommendations
42+
43+
### Suggested Order
44+
45+
1. **Math foundations**: Linear algebra, probability theory, calculus
46+
2. **Machine learning**: Understanding classical ML algorithms
47+
3. **Deep learning**: Neural network basics and backpropagation
48+
4. **Modern architectures**: Transformer and attention mechanisms
49+
5. **Applied practice**: Applying models to specific tasks
50+
51+
### Practical Tips
52+
53+
1. **Balance theory and practice**: Implement every concept you learn
54+
2. **Project-driven**: Consolidate knowledge through complete projects
55+
3. **Community participation**: Join learning communities for discussion
56+
4. **Stay current**: Keep up with the latest technical developments
57+
58+
### Common Challenges
59+
60+
1. **Math barrier**: Requires some mathematical background
61+
2. **Abstract concepts**: Some ideas are abstract and require hands-on practice
62+
3. **Fast-moving field**: Requires continuous learning of new techniques
63+
4. **Theory-practice balance**: Balancing theoretical study with practical work
64+
65+
## Advanced Directions
66+
67+
### Theoretical Deepening
68+
69+
- Optimization theory and algorithms
70+
- Information theory and deep learning
71+
- Statistical learning theory
72+
- Bayesian deep learning
73+
74+
### Application Domains
75+
76+
- Computer vision
77+
- Natural language processing
78+
- Speech recognition and synthesis
79+
- Recommender systems
80+
81+
### Engineering Practice
82+
83+
- Large-scale training
84+
- Model deployment and optimization
85+
- Distributed computing
86+
- MLOps practices
87+
88+
## Resource Summary
89+
90+
### Online Courses
91+
92+
- MIT 6.034 Artificial Intelligence
93+
- Stanford CS229 Machine Learning
94+
- Deep Learning Specialization (Coursera)
95+
- Fast.ai Practical Deep Learning
96+
97+
### Classic Textbooks
98+
99+
- _Deep Learning_ (Goodfellow et al., the "Bible")
100+
- _Machine Learning_ (Zhihua Zhou, the "Watermelon Book")
101+
- _Statistical Learning Methods_
102+
- _Pattern Recognition and Machine Learning_
103+
104+
### Practice Platforms
105+
106+
- Kaggle competition platform
107+
- Google Colab
108+
- Jupyter Notebook
109+
- GitHub open-source projects
110+
111+
These resources provide a complete learning path from theory to practice in deep learning. Choose the approach that best suits your background and goals.

‎app/docs/ai/llm-basics/deep-learning/misc/index.en.mdx‎

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
---
2+
title: Other Resources
3+
description: Miscellaneous deep learning resources.
4+
docId: lodydcd211esraq1r55ze9ey
5+
lang: en
6+
translatedFrom: zh
7+
translatedAt: 2026-04-15T12:00:00Z
8+
translatorAgent: claude-sonnet-4-6
9+
---
10+
11+
# Other Important Resources
12+
13+
## Machine Learning "Pumpkin Book"
14+
15+
- **Core value**: Detailed derivations of the formulas in Zhihua Zhou's _Machine Learning_
16+
- **Target audience**: Learners who want a deep understanding of the mathematical principles behind algorithms
17+
- **Highlights**: Thorough mathematical derivations and proofs
18+
19+
## Li Hongyi Deep Learning Notes
20+
21+
- **GitHub repo**: https://github.com/datawhalechina/leedl-tutorial
22+
- **Core value**: Very low barrier to entry; covers virtually every aspect of deep learning
23+
- **Target audience**: Can be read directly even without prior machine learning knowledge
24+
- **Highlights**: Friendly to readers with weaker math backgrounds
25+
26+
## The "Bible" of Deep Learning
27+
28+
- **Authors**: Ian Goodfellow, Yoshua Bengio, Aaron Courville
29+
- **Highlights**: Solid theoretical foundations, mathematically rigorous
30+
- **Coverage**: Theoretical foundations and mathematical principles of deep learning
31+
- **Suitable for**: Researchers and advanced practitioners
32+
33+
## MIT Deep Learning Textbook
34+
35+
- **Official website**: [https://www.deeplearningbook.org/](https://www.deeplearningbook.org/)
36+
- **Highlights**: Authoritative theoretical reference, academic standard
37+
- **Language**: English original, high theoretical depth
38+
39+
## PKU CSDIY
40+
41+
- **Resource link**: [https://csdiy.wiki/](https://csdiy.wiki/)
42+
- **Highlights**: Self-study guide for computer science, includes a deep learning path
43+
- **Coverage**: Course recommendations, learning roadmaps, hands-on projects
44+
45+
## Video Courses
46+
47+
### Gupao Programmer AI Course
48+
49+
- **Course**: [AI & Machine Learning] 2023 Comprehensive System Tutorial
50+
- **Content**: Machine learning algorithms, machine learning in practice
51+
- **Link**: [Bilibili Video](https://www.bilibili.com/video/BV1hM4y1U7FV)
52+
- **Highlights**: Systematic coverage, practice-oriented

‎app/docs/ai/llm-basics/deep-learning/nlp/index.en.mdx‎

Lines changed: 24 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
---
2+
title: NLP
3+
description: Natural language processing fundamentals.
4+
docId: nrelvvfzq0gma7pqfx9fkfxt
5+
lang: en
6+
translatedFrom: zh
7+
translatedAt: 2026-04-15T12:00:00Z
8+
translatorAgent: claude-sonnet-4-6
9+
---
10+
11+
# NLP Fundamentals
12+
13+
## HuggingFace NLP Course
14+
15+
- **Official link**: [https://huggingface.co/learn/llm-course/zh-CN/chapter7/2](https://huggingface.co/learn/llm-course/zh-CN/chapter7/2)
16+
- **Highlights**: Practical NLP tools and techniques
17+
- **Content**: Using the Transformers library, fine-tuning models, deployment, and more
18+
19+
## Learning Path
20+
21+
1. **Fundamentals**: Text preprocessing, word vectors, language models
22+
2. **Classic models**: RNN, LSTM, GRU, Attention
23+
3. **Modern architectures**: Transformer, BERT, GPT series
24+
4. **Applied practice**: Text classification, named entity recognition, machine translation

‎app/docs/ai/llm-basics/embeddings/index.en.mdx‎

Lines changed: 43 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,43 @@
1+
---
2+
title: Embedding Models
3+
description: Resources on embeddings and vector representations.
4+
date: "2024-01-15"
5+
tags:
6+
- embedding
7+
- vector
8+
- representation
9+
- nlp
10+
- retrieval
11+
docId: xnl2yzrb4x748zhhfe26ragt
12+
lang: en
13+
translatedFrom: zh
14+
translatedAt: 2026-04-15T12:00:00Z
15+
translatorAgent: claude-sonnet-4-6
16+
---
17+
18+
**Embedding** is a technique that maps discrete objects (such as words, sentences, images, and user behaviors) to a continuous vector space.
19+
With this representation, semantically similar objects tend to be closer together in the vector space, making computation and modeling more tractable.
20+
21+
## Core Idea
22+
23+
- **Discrete → Continuous**: Transforms symbolic inputs into numerical vectors, enabling neural network processing.
24+
- **Semantic preservation**: The structure of the vector space retains the semantic relationships between objects.
25+
- **Computability**: Vectors support operations such as addition, dot product, and cosine similarity, enabling retrieval, clustering, and classification.
26+
27+
## Applications in Large Models
28+
29+
- **Word/sentence vectors**: The most common representation in NLP models (e.g., Word2Vec, BERT, GPT).
30+
- **Multimodal representations**: Mapping images, audio, video, and other modalities into a shared vector space for cross-modal retrieval.
31+
- **Retrieval and recommendation**: Semantic retrieval based on vector similarity (vector databases, RAG), and personalized recommender systems.
32+
- **Fine-tuning and merging**: Optimizing vector representations for specific tasks via methods such as LoRA and SLERP.
33+
34+
## Typical Methods
35+
36+
- **Early methods**: Word2Vec, GloVe
37+
- **Contextual representations**: ELMo, BERT
38+
- **Embeddings from generative LLMs**: GPT series, Qwen Embedding, OpenAI Embedding API
39+
40+
## Summary
41+
42+
Embedding is a foundational component of modern machine learning and large model applications.
43+
It bridges the discrete and continuous worlds, and is a core tool for semantic understanding, retrieval-augmented generation (RAG), and multimodal fusion.

0 commit comments

Comments
 (0)