hahhforest
Featured
pi-textbook
A hands-on course that walks through building a Pi-style coding agent from scratch across 15 checkpoints, starting from an offline agent trajectory.
from-minimind-to-more-notes
This repository provides structured reading notes for 17 articles (approx. 620k words) from the 'from-minimind-to-more' project, covering LLM foundations, architectures, algorithms (Pretrain, SFT, DPO, PPO, GRPO, SPO), and interview preparation.
Stars
7
7-day growth
No data
Forks
0
Open issues
0
License
NOASSERTION
Last updated
2026-07-29
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It offers a meticulously organized, multi-angle summary of a substantial educational resource on LLMs, complete with critical annotations, multiple reading paths, and code-level references to the minimalist minimind implementation.
Who it is for
- Students and self-learners of deep learning and LLMs
- AI job seekers preparing for technical interviews
- Researchers or engineers seeking a structured overview of modern LLM training pipelines
- Educators looking for curated teaching materials on RLHF and transformer variants
Use cases
- Self-paced study of LLM internals from tokenization to RLHF
- Interview prep for machine learning engineering roles
- Cross-referencing minimind code with theoretical explanations
- Quick revision of key concepts, formulas, and algorithms in LLM development
Strengths
- Consistent 9-section note structure for every article, including code analysis and critical notes
- Multiple curated reading paths tailored to different backgrounds and goals
- Covers both foundational theory and practical algorithm details (e.g., KV Cache, MoE, PPO/GRPO)
- Includes explicit 'critical annotations' that note gaps between teaching implementation and industrial practice
Considerations
- Repository contains only reading notes, not the original articles or code
- Some content (e.g., Chinchilla scaling laws, classic RLHF narrative) may be outdated as of 2026
- Based on a minimal teaching implementation (minimind) that diverges from production-grade systems (e.g., one-step MDP in PPO, mean instead of sum in DPO)
README quick start
From Minimind to More — 精读结构化笔记
对 Tongyun1/from-minimind-to-more 全 17 篇正文(约 62 万字 / 1 万行)的逐篇精读合成笔记。
特别致谢:本笔记基于 Tongyun1 的原项目整理,原项目又基于 minimind(作者 jingyaogong)与 MiniMind-in-Depth。感谢原作者无私开源。本仓库仅含精读笔记,不含原正文,所有版权归原项目作者所有。
这是什么
一份对 from-minimind-to-more 仓库的逐篇结构化精读笔记,覆盖:
- 基石篇(3):Tokenizer、Minimind 设计目录、Embedding 与位置编码
- 架构篇(5):归一化技术、KV Cache→Flash Attention、MoE、超级拼装、推理训练优化
- 算法篇(7):RL 概览、Pretrain、SFT、DPO、PPO、GRPO 及变体、SPO
- 求职篇(1):大模型八股 100 问
笔记结构
每篇笔记含固定八段:
- 一句话精炼
- 核心概念(分点列出)
- 关键公式(LaTeX)
- 关键算法/流程(伪代码或步骤)
- 源码要点(对照 minimind 代码逐行)
- 作者独到见解/类比(摘录原文)
- 面试考点
- 批判性批注(事实性/覆盖面/过时风险)
- 篇内思维导图(mermaid 或缩进树)
文件清单
| 文件 | 内容 |
|---|---|
00_OVERVIEW.md | 总图 + 总索引(全篇知识 mindmap、训练链路、GRPO 家族谱系、工程速查表、全局批判、阅读路径) |
01–16_*.md | 16 篇正文逐篇精读笔记 |
建议先读 00_OVERVIEW.md 建立全貌,再按其中推荐的阅读路径精读单篇。
阅读路径
- 新手:
02 → 01 → 03 → 04 → 07 → 10 → 11 → 09 → 16 - 求职冲刺:
16 → 04/05/06 → 09 → 12/13/14 → 速查表 - 算法岗:
09 → 12 → 14 → 15 → 13 → 10/11 - 工程/推理岗:
05 → 08 → 06 → 07 → 速查表
免责声明
- 本笔记是个人学习产物,不代表原项目作者观点。
- minimind 是极简教学实现,与工业标准有差距(如 PPO 用 One-step MDP、DPO 用 mean 而非 sum 等),笔记在"批判性批注"段已逐项标注。
- 部分内容(Chinchilla scaling law、经典 RLHF 叙事、MMLU 等)在 2026 视角下已过时,请带"时间戳意识"交叉验证。
- 遇关键论断建议回原论文核对。
License
笔记内容遵循原项目精神开源。原项目 from-minimind-to-more 未声明 License,版权归原作者所有。
Description
对 Tongyun1/from-minimind-to-more 的逐篇结构化精读笔记(基石/架构/算法/求职,含思维导图与批判性批注)
Related repositories
Similar projects matched by category, topics, and programming language.
makecindy
Featured
cindy
Cindy is an open-source AI agent that runs locally on your machine, integrates multiple AI harnesses and models, and provides memory, skills, and automation to perform real work in your projects and apps.
uzairansaruzi
Featured
hermex
Hermex is a native SwiftUI iPhone app that lets you control a self-hosted Hermes AI agent directly from your phone, with no subscriptions, tracking, or third-party relay.