中文内容在规划中了，请加星关注（便于收到更新通知）
A professional, four-track curriculum covering the full lifecycle of Large Language Models — from linguistic foundations to production deployment.
Track 1 Track 2 Track 3 Track 4
Fundamentals Scientist Engineering Solutions
(NLP → Transformer) (Research & Training) (Build & Operate) (Architect & Ship)
│ │ │ │
▼ ▼ ▼ ▼
Prerequisite ──────→ Deep Theory ──────→ Industrial Practice ──────→ Business Delivery
From Classical NLP to the Attention Revolution.
| Module | Core Topics | Entry Point |
|---|---|---|
| 01 Linguistics | NLP hierarchy, morphology, syntax, semantics | Linguistic Foundations |
| 02 Classical NLP | Preprocessing, BoW/TF-IDF, HMM/CRF | Text Preprocessing |
| 03 Deep Learning | Word2Vec/GloVe, RNN/LSTM/GRU, Seq2Seq | Word Embeddings, Word2Vec Demo |
| 04 Transformer Era | Attention mechanism, Transformer architecture, Pre-train paradigms | Attention, Attention Viz |
| 05 Applications | Classification, NER, MT, Summarization, Dialogue, Search | LLM Disruption Map |
State-of-the-art model architecture, training, alignment, and frontier research.
| Module | Core Topics | Entry Point |
|---|---|---|
| 01 Architecture (12) | Transformer, MHA/MQA/GQA/MLA, Efficient Attention, Tokenizer, Embedding, RoPE, Dense vs MoE, Decoding, Interpretability, Long Context | Transformer |
| 02 Dataset (5) | Pre-training data at scale, Instruction data, Preference data, Synthetic data, PII management | Data at Scale |
| 03 Pre-Training (11) | GPT evolution, Scaling Laws, Attention optimizations, Data pipelines, Distributed training, Stability, Continual pre-training | Scaling Laws |
| 04 Post-Training | FT: PEFT/LoRA/QLoRA, Domain adaptation | PEFT Strategies |
| Alignment: PPO, DPO, KTO, RLAIF, Constitutional AI, RLVR, GRPO | Alignment Overview | |
| Advanced: Rejection Sampling, Iterative Training, Inference-Time Compute, Model Merging | Inference-Time Compute | |
| Distillation | Distillation Overview | |
| 05 Evaluation (5) | Benchmarks taxonomy, Methodology, LLM-as-Judge, Safety eval, Contamination detection | Benchmarks |
| 06 Multimodal (4) | Vision-Language, Audio/Speech, Video understanding, Multimodal eval | VLM |
| 07 Paper Tracking (5) | Tracking methodology, Architecture/Training/Alignment/Multimodal frontiers | Methodology |
Building, deploying, and operating production-grade LLM applications.
Every module follows a strict 3-layer structure: Theory → Practical (.py) → Best Practice (.md).
| Module | Theory | Practical | Best Practice |
|---|---|---|---|
| 01 LLMs | Intelligence landscape, Tokenization & cost, API mechanics, Engineering paradigm | Async Gateway, Batch API, Guardrails | Architecture Matrix, Model Routing |
| 02 Prompt Engineering | Foundations, Programmatic prompting, Reasoning strategies, Structured Output & Function Calling, Prompt Template Architecture, Data-Driven Prompt Design | DSPy, Self-Correction, Structured Output | Prompt CI/CD, Defensive Design |
| 03 Context Engineering | Context window mechanics, Context composition & priority, Token budget & cost, Long context techniques, Dynamic context management, Advanced paradigms, CE Evaluation | Shared: Composer, Budget Controller, Compressor, Observability · Cases: Customer Support, Document Analysis | Architecture Patterns, Quality & Eval, Production Optimization, Vendor Practices |
| 04 Memory | Memory systems, Cross-session persistence | Sliding Window, Vector Memory | Architecture Patterns |
| 05 RAG | Architecture, Advanced RAG, Data ingestion, GraphRAG | Query Routing, Hybrid Indexing, Reranking | RAG Eval Framework, Embedding Selection |
| 06 Agent | Theory, Architecture, Workflow patterns, Multi-agent, MCP Protocol | ReAct Agent, Multi-Agent, MCP Server | Agent Eval, Production Guardrails |
| Frameworks (9): ADK, CrewAI, CamelAI, Agno, LangGraph, AutoGPT, BabyAGI, Semantic Kernel, OpenAI Swarm | ADK Agent, Agno Agent | ||
| 07 Deployment | Optimization, Architecture, Quantization, Cloud comparison | vLLM, Continuous Batching | Production Checklist, SLOs & Monitoring |
| 08 Security | LLM threats, Advanced threat modeling, Privacy/Compliance, Secure architecture | Injection Detection, PII Redaction, Agent Sandbox | Compliance Checklist, Incident Response |
| 09 LLMOps | Maintenance, Observability, CI/CD for LLMs | Eval Runner, Observability Collector | Production Checklist, On-Call Runbook |
Architectural decision frameworks and implementation roadmaps for domain LLM applications.
Four-phase progression: Strategy → Infrastructure → Build → Ship.
| Phase | Document | Key Question |
|---|---|---|
| Strategy | 01 Technology Selection | Prompt Eng vs RAG vs Fine-tuning? |
| 02 Cost & ROI Analysis | Is it worth building? | |
| Infrastructure | 03 Domain Data Strategy | Where does the data come from? |
| 04 Evaluation Loop | How do we measure success? | |
| Build | 05 RAG Architecture | Multi-source, Agentic RAG patterns. |
| 06 Finetuning Playbook | CPT → SFT → DPO execution guide. | |
| 07 Knowledge Graph Integration | Hybrid structured + unstructured. | |
| 08 Agent Workflow Design | Business process orchestration. | |
| Ship | 09 Vertical Scenario Templates | Legal, Finance, Manufacturing, Medical blueprints. |
| 10 Implementation Roadmap | PoC → MVP → Production → Scale. |
| Goal | Recommended Path |
|---|---|
| "I'm new to NLP/LLM" | Track 1 (all) → Track 3 (01-02) → Track 4 (01) |
| "I want to build LLM apps" | Track 3 (01→09) → Track 4 (01→10) |
| "I want to train/align models" | Track 1 (04) → Track 2 (01→05) |
| "I need to deploy to production" | Track 3 (07→09) → Track 4 (02, 10) |
| "I'm evaluating LLM for my business" | Track 4 (01→02) → Track 3 (01) → Track 4 (10) |
| Type | Title | Tracks |
|---|---|---|
| Book | Build a Large Language Model — Raschka (2025) (materials/books/) |
Track 1–2: end-to-end guide from architecture → pre-training → fine-tuning → alignment |