wdndev/llm_interview_note: a Chinese-language LLM interview question bank you read in the browser
主要记录大语言大模型（LLMs） 算法（应用）工程师相关的知识及面试题
At a glance
- What is it?
- A Docsify site and Markdown repository covering LLM fundamentals, training, inference and RAG, aimed at Chinese-speaking algorithm engineers preparing for interviews. The content is one author's own notes, and the repository does not state a licence.
- Who is it for?
- Use it if you read Chinese and want one place to revise LLM fundamentals, distributed training, SFT, inference and RAG before an interview, and you accept that answers are one person's notes rather than a reviewed reference. Do not adopt it as a source of record for production decisions or as a substitute for the original papers, and do not expect an English translation or a stated licence.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 107 days ago.
- What is it written in?
- Mainly HTML, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 29, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What llm_interview_note actually is, and who it is written for
The repository is a collection of Markdown notes on large language model concepts, organised as interview preparation material for algorithm and application engineers. The README opens by describing the repository as large-model interview concepts compiled from online resources by the author, and it asks readers to point out anything unreasonable. That framing matters: this is a personal study record published openly, not a curated textbook and not a vendor knowledge base.
The audience is narrow and explicit. Everything is written in Chinese, and the topics list assumes you already know what a transformer is and want to be questioned on it. Sections run from language model basics and tokenisation through transformer architecture, MoE, distributed training, supervised fine-tuning, inference frameworks, reinforcement learning, RAG, and evaluation. There is also a separate chapter of real interview questions and a folder of related courses and references.
The author maintains a set of companion hands-on repositories listed in the README: tiny-llm-zh for building a small Chinese language model from scratch, tiny-rag for a retrieval system with multi-way recall and reranking, tiny-mcp for an MCP server and client using prompts and function calling, and llama3-from-scratch-zh, which the README says can load Meta's official weights and run on a local laptop with 16 GB of memory. Those are separate projects; this repository is the reading material, not the code.
One structural detail is worth knowing before you start reading. The primary language reported for the repository is HTML, because the site is rendered by Docsify: index.html loads the Markdown files into the browser at read time. The Markdown under the numbered directories is the real content.
How the Docsify site and the Markdown tree fit together
There is no build step and no static site generator producing pre-rendered pages. The repository root holds index.html, _navbar.md, _sidebar.md and a .nojekyll file, which is the standard combination for Docsify on GitHub Pages. The .nojekyll file stops GitHub Pages from running Jekyll over the output, so directories whose names begin with digits and contain Chinese characters are served as-is. The navigation bar and the sidebar are themselves Markdown files, which means the table of contents in the README and the sidebar you click in the browser are two copies of the same list and can drift apart.
When a reader opens a page, the browser fetches the corresponding .md file and renders it client side. The practical consequence is that a page cannot be read until its Markdown file has been downloaded, and search engines see whatever the renderer produces rather than server-rendered HTML. For a documentation site whose whole purpose is reading, that trade is acceptable, and it is the reason the repository can be forked and served from any static host with the same layout.
The directory names encode the reading order: 01.大语言模型基础, 02.大语言模型架构, 03.训练数据集, 04.分布式训练, 05.有监督微调, 06.推理, 07.强化学习, 08.检索增强rag, 09.大语言模型评估, 10.大语言模型应用, plus 98.相关课程 and 99.参考资料 for external material. Inside each chapter, individual topics are their own directories with a single Markdown file of the same name, so a link such as 02.大语言模型架构/1.attention/1.attention.md resolves directly. That one-file-per-topic layout is what makes the repository easy to read on GitHub without the site at all.
A pdf_note directory also exists at the root. The README does not describe what it contains or how it is generated, so treat any PDF you find there as an artefact of unknown freshness rather than a released edition.
Reading it online or cloning it locally
The README gives one official reading link, the GitHub Pages site. There is no package to install and no CLI, so the only decision is whether you read it in the browser or keep a local copy.
The README does not document a clone command, a local preview command, a package manager or a build script, so there is nothing to install for the site itself. What the repository does give you is a plain directory tree of Markdown files, and the README's table of contents maps every topic to its file path. If you want an offline copy, use whatever Git client you already have against the repository URL shown on GitHub, then open the numbered chapter directories in a Markdown viewer.
The README does not state where topic images are stored. The one image it embeds, the author's WeChat account QR code, is hosted in a separate personal repository, so a local copy can render text correctly while some images fail to load. Check that before assuming a clone is complete.
If you only want the interview questions rather than the theory, the README's table of contents points to a chapter titled 真实面试题, and the later chapters each end with a section of practice questions, for example 分布式训练题目, 微调, 推理 and rlhf相关. Those are the pages to bookmark if your preparation window is short.
Where the notes are thin, and when this is the wrong tool
The README is direct about provenance: the answers are written by the author, drawing on online resources, and corrections are invited. That is a reasonable way to publish study notes and a poor basis for anything you need to be correct. There is no review process described, no list of contributors, and no citation apparatus beyond the 99.参考资料 directory of references. If an answer conflicts with a paper, the paper wins.
Coverage is uneven by design. Some chapters are fully written, while others are placeholders in the table of contents with no linked file: 03.训练数据集 has a 3.2 模型参数 entry with no page, 04.分布式训练 lists 4.3 Megatron, 4.4 训练加速 and 4.5 一些有用的文章 as empty headings, 06.推理 has 6.3 量化 and 6.4 vLLM as bare entries, and 07.强化学习 has a 7.3 一些题目 heading without the sub-items spelled out in the README. A reader who works through the sidebar top to bottom will hit gaps that the table of contents does not signal.
The repository also has no stated licence. Nothing in the README grants reuse terms, so copying chapters into internal training material, a paid course, or a product knowledge base is a decision made without a grant. That is a real constraint for anyone at a company, not a formality.
Finally, the language. There is no English edition and the README does not mention a translation. If your interview will be conducted in English, or if you need to quote terminology in English under pressure, this repository trains the wrong reflex. It is also the wrong tool if what you need is runnable code: the hands-on work lives in the author's separate tiny-llm-zh, tiny-rag and tiny-mcp repositories, and this one is prose.
How it compares with a general machine learning interview repository
The closest alternative in the author's own list is ai_interview_note, described in the README as covering deep learning, machine learning, recommender systems and search systems as general knowledge. The difference in approach is scope, not depth. ai_interview_note spreads across the broader AI engineering surface, while llm_interview_note concentrates on the large language model stack: architecture choices such as decoder-only versus encoder-decoder, attention variants like MHA, MQA and GQA, decoding strategies including top-k, top-p and temperature, parallelism strategies, LoRA and adapter tuning, inference frameworks, RLHF with PPO and DPO, and RAG.
Compared with a paper-reading habit, the trade is inverted. Papers give you the primary argument and the experiment, and take an afternoon each. This repository gives you a condensed Chinese summary in a few minutes, at the cost of losing the evidence and the caveats. For interview preparation the compression is the point; for engineering decisions it is a liability.
Compared with a question-and-answer site, the difference is that the questions here are grouped by topic with the surrounding theory in the same repository, so you can read the concept and then test yourself on it without switching sources. The README does not claim the real interview questions are statistically representative of anything, and they should not be treated as a syllabus.
Maintenance, upgrades and licence status
The repository is not archived, and the last push was on 2026-06-14. That is roughly three months before today, which puts it inside the six-month window, but the README gives no release schedule and there are no releases. Updates arrive as ordinary commits, and the README says the author also posts LLM content and interview experience on a WeChat account on an irregular basis, which is the only cadence statement available.
Upgrade cost is close to zero in the mechanical sense. There is no dependency to bump and no build to rerun: you pull the repository, or you reload the site, and you have the current text. The real cost is editorial. Because the content tracks a fast-moving field, chapters on inference frameworks and reinforcement learning can age faster than chapters on tokenisation, and nothing in the repository marks a page as stale. If you keep a local clone, the only way to know whether a page changed is to look at its commit history on GitHub, which is one reason to read the files on the platform rather than only on the rendered site.
On licensing, the repository metadata does not identify a licence and the README does not state one. Without a licence, the default position under copyright is that no reuse rights are granted, so redistribution and commercial reuse are unclear. This is a description of the repository's own state, not legal advice; if reuse matters to you, ask the author through the channels the README lists.
Editorial conclusion
Use it if you read Chinese and want one place to revise LLM fundamentals, distributed training, SFT, inference and RAG before an interview, and you accept that answers are one person's notes rather than a reviewed reference. Do not adopt it as a source of record for production decisions or as a substitute for the original papers, and do not expect an English translation or a stated licence. Before relying on any page, open the Markdown file on GitHub, check the last commit date on that file, and confirm the repository still carries no LICENSE file.
Frequently asked questions
What are some common questions related to LLMs?
The repository groups its material by topic rather than listing questions alone: language model basics, tokenisation and embeddings, transformer architecture including attention and layer normalisation, MoE, distributed training, supervised fine-tuning, inference frameworks, reinforcement learning, RAG and evaluation. Each of the later chapters ends with a practice question section, and a separate chapter collects real interview questions.
What is LLM in brief?
The repository does not open with a one-line definition; it starts from language models and builds up through tokenisation, word vectors and the transformer. If you want a short definition, this is not the page for it, and the README points instead to the author's general ai_interview_note repository for broader background.
What are some common interview questions about fine tuning in LLM programs?
Chapter 05 covers supervised fine-tuning, with pages on basic concepts, prompting, adapter-tuning and LoRA, plus practical walkthroughs for fine-tuning llama2 and ChatGLM3. It closes with practice question pages on fine-tuning and pretraining.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/wdndev-llm-interview-note)
Community notes