LLM course · LLM roadmap · large language models · fine-tuning · RAG · agents · quantization · evaluation · serving
Start here · Fundamentals · Scientist · Engineer · Operator · Leader · Labs · YouTube · Glossary · Why this exists
Developed by Zorost Intelligence AI Lab · Washington, DC · Training 02 on zorost.com/ai-lab
LLM Engineering is a free, original LLM course and LLM roadmap for people who work with large language models. It teaches how models are built, how they are adapted (fine-tuning, LoRA, DPO), how products use them (RAG, agents, structured output), how they are served, and how a technical leader decides what to fund.
It is original writing, original hub-and-leaf roadmaps, and original CPU-first labs from Zorost Intelligence AI Lab. MIT licensed. No signup. The required labs run on a laptop CPU with no API key.
It sits next to AI Engineering Lab (Training 01), which takes a beginner from Python to a governed production lakehouse in 24 weeks. This repository is the specialist map of the LLM field itself.
The job is not "know the names of papers." The job is to choose a path, measure the result, and know why the other path would have failed.
| If you are… | Open this first | Then do |
|---|---|---|
| New to machine learning | START-HERE then Fundamentals | Labs 00 to 03 |
| Training or adapting models | Scientist | Labs 04 to 08, 17 to 19 |
| Shipping products on models | Engineer | Labs 08 to 16 |
| Running models in production | Operator | Labs 13, 15, 18 |
| Deciding budget, risk, and org | Leader | The four leader modules, then the textbooks |
Do not start at agents if you cannot yet explain a token, a context window, and an eval. The decision diagram is the short version of that rule.
git clone https://github.com/zorost/LLM-Engineering.git
cd LLM-Engineering
python3 -m pip install -r requirements.txt
python3 scripts/check.pyThen open START-HERE.md. If you already write Python and have trained a network, skip to the track that matches your job.
The vocabulary under every later sentence. Five hubs. Hub 05 is the hardware sitting public LLM courses usually omit: scaling laws, VRAM, KV cache, MoE, token cost.
| # | Module | What you can do after it |
|---|---|---|
| F1 | Mathematics | Read a gradient, a matrix multiply, and a probability statement without bluffing |
| F2 | Python and data | Load, split, and plot a dataset, and know why leakage is fatal |
| F3 | Neural networks | Train a tiny network and name the failure (underfit, overfit, exploding) |
| F4 | NLP before transformers | Explain why bag-of-words dies on meaning, and what embeddings fixed |
| F5 | Scaling, information, hardware | Estimate tokens, FLOPs, and VRAM before you buy a box |
Track README · Pipeline diagram
How capable models are made: architecture, pre-training, post-training data, supervised fine-tuning, preference alignment, evaluation, quantization. Nine hubs. Reasoning, distillation, and licenses are first-class modules, not a single "new trends" column.
You do not need a GPU cluster to understand this track. You do need one if you intend to reproduce frontier pre-training.
| # | Module | What you can do after it |
|---|---|---|
| S1 | Architecture | Trace text through tokenize, embed, attend, sample; know RoPE, GQA, MoE |
| S2 | Pre-training | Describe data, parallelism, and the metrics that say a run is dying |
| S3 | Post-training data | Design an instruction or preference set, and filter it |
| S4 | Supervised fine-tuning | Choose full FT vs LoRA vs QLoRA and name the parameters that matter |
| S5 | Preference alignment | Contrast DPO, PPO, GRPO, ORPO, and when RL is worth the cost |
| S6 | Evaluation | Build a harness that is not just a leaderboard screenshot |
| S7 | Quantization | Pick GGUF, GPTQ, AWQ, or bitsandbytes for a real constraint |
| S8 | Merging, multimodal, interp | Merge, ablate, or condition on images without cargo-culting the tool |
| S9 | Reasoning and test-time | Scale compute at decode time on purpose, not by accident |
| S10 | Distillation and draft models | Explain speculative decoding and why a draft model helps |
| S11 | Licenses, contamination, data law | Read a model card and a dataset card before you ship |
| S12 | Interpretability and editing | Know what SAEs and activation steering can and cannot claim |
How applications that use models are built. This is the default working path for most readers of this LLM course. Retrieval before agents. Security and evals before the demo.
| # | Module | What you can do after it |
|---|---|---|
| E1 | Running models | Call an API or a local runtime, and constrain the output |
| E2 | Vector storage | Ingest, chunk, embed, and retrieve with a metric |
| E3 | RAG | Ground an answer, cite it, and measure faithfulness |
| E4 | Advanced RAG | Hybrid search, rerank, graphs, text-to-SQL, query rewrite |
| E5 | Agents | Build a tool loop before you import a framework |
| E6 | Inference optimization | KV cache, Flash Attention, batching, speculative decode |
| E7 | Deployment | Local, demo, server, and edge, with an honest cost line |
| E8 | Security | Prompt injection, poisoning, red teaming, logging of secrets |
| E9 | Context, tools, MCP | Pack a window, cache a prefix, expose a tool through MCP |
| E10 | Evals in production | Gate a release on a suite, not a vibe |
| E11 | Memory and conversation | Buffer, summarize, retrieve; know when each lies |
| E12 | Platforms | Azure, Vertex, Bedrock, Databricks, and open gateways |
The track most public LLM roadmaps skip. The scientist trains. The engineer ships a product. Someone still has to keep a model answering under a latency budget, on a GPU bill that finance can defend.
| # | Module | What you can do after it |
|---|---|---|
| O1 | Serving engines | Choose vLLM, SGLang, TGI, llama.cpp, or MLX for a workload |
| O2 | Reliability | Timeouts, retries, fallbacks, circuit breakers, idempotency |
| O3 | Cost and capacity | Tokens, QPS, VRAM, and the bill |
| O4 | Observability | Traces, eval samples, and privacy-safe logs |
For people who allocate money, risk, and attention. Written against The AI Leadership Textbook and AI Engineering Distilled (author page).
| # | Module | What you can do after it |
|---|---|---|
| L1 | Buy, build, retrieve, or fine-tune | Pick the cheapest move that hits the eval |
| L2 | Teams and skills | Staff an LLM effort without cloning a research lab |
| L3 | Risk and governance | Map NIST AI RMF and the EU AI Act onto a real system |
| L4 | Portfolio and product | Kill a demo that will never become a product |
Each poster is a spine of numbered hubs. A hub is a module family. The boxes on the opposite side are leaves: topics inside that hub, not extra courses. The lava hub is the sitting people skip or start too late.
HTML sources live in assets/diagrams/. The README copies
live in img/. Original Zorost artwork. Orthogonal connectors, PAPERG
paper, INK type, LAVA accent. We kept the coverage shape of the public
LLM-course tradition (fundamentals, scientist, engineer) and rewrote the
pictures. How the posters were designed.
Twenty notebooks. All of them run on CPU. None of them require a paid key. They exist so you can break the idea, not so you can screenshot a Colab GPU.
Notebook index · Tutorial notes
| Lab | Notebook | Concept |
|---|---|---|
| 00 | 00_environment_check.ipynb |
Install, versions, the check script |
| 01 | 01_tokenization_bpe.ipynb |
Byte-pair encoding from scratch |
| 02 | 02_attention.ipynb |
One-head attention on tiny tensors |
| 03 | 03_decoding_strategies.ipynb |
Greedy, temperature, top-k, nucleus |
| 04 | 04_chat_templates.ipynb |
Why a missing end token wrecks SFT |
| 05 | 05_lora_shapes.ipynb |
LoRA rank, alpha, and parameter count |
| 06 | 06_dpo_loss.ipynb |
Direct preference optimization on toy pairs |
| 07 | 07_quantization.ipynb |
Absmax and zero-point, with error |
| 08 | 08_embeddings_rag.ipynb |
Chunk, embed, retrieve, cite |
| 09 | 09_hybrid_retrieval.ipynb |
Sparse plus dense, then rerank |
| 10 | 10_react_agent.ipynb |
Thought, action, observation, without a framework |
| 11 | 11_structured_output.ipynb |
JSON schema as a contract |
| 12 | 12_eval_harness.ipynb |
A suite, a gate, an error bucket |
| 13 | 13_vram_and_cost.ipynb |
Memory math and token economics |
| 14 | 14_prompt_injection_defense.ipynb |
Attacks as test cases, not as a cookbook |
| 15 | 15_openai_compatible_client.ipynb |
One client, many backends |
| 16 | 16_speculative_decoding.ipynb |
Draft and verify on a toy alphabet |
| 17 | 17_moe_routing.ipynb |
Experts, gates, load balance |
| 18 | 18_kv_cache.ipynb |
Why the second token is cheaper |
| 19 | 19_scaling_laws.ipynb |
Loss vs compute, with honest limits |
Run one:
python3 -m jupyter notebook notebooks/00_environment_check.ipynbOr execute them all headless after you trust scripts/check.py.
This repository is one shelf in a public set. Use the others. Do not copy client work into either.
| What | Where | Use it for |
|---|---|---|
| AI Engineering Lab (Training 01) | github.com/zorost/AI-Engineering-Lab · zorost.com/ai-engineering-lab | 24-week path from Python to Databricks, with a single case study |
| Transformer Explainer | github.com/zorost/transformer-explainer | In-browser attention, RoPE, GQA, KV cache, MoE |
| 30 days of Databricks | github.com/zorost/30-days-of-Databricks | Lakehouse literacy that the Engineer platforms module assumes |
| AI Fieldwork | zorost.com/ai-lab/fieldwork | Dated experiments: what held, what broke |
| Open tools | github.com/zorost · zorost.com/ai-lab/open-source | DocPrep-AI, MarkForge, yt2textbook, OneRead, SparkDuet, AlchemyLake |
| The AI Leadership Textbook | Amazon · Author page | Silicon to boardroom: 44 chapters, platforms, governance, playbooks |
| AI Engineering Distilled | Amazon | Seam notation for specifying LLM, RAG, and agent systems |
Full citations live in reference/BOOKS.md. The public training register is zorost.com/ai-lab/training.
A curated, dated map of lectures we actually send people to. Not a dump of every thumbnail.
Start with these, in this order:
- 3Blue1Brown, But what is a neural network? then the Transformer
- Andrej Karpathy, Deep Dive into LLMs like ChatGPT then Let's build GPT and Let's build a tokenizer
- Stanford CS336, Language Modeling from Scratch (2026) (2025 playlist, course site)
- Hugging Face NLP Course then LLM Course and Agents Course
- fast.ai Practical Deep Learning
The full list, with why each item is there, is reference/YOUTUBE.md. Outside courses we respect and do not copy: reference/COURSES.md.
LLM-Engineering/
├── START-HERE.md # day one: install, order, when it breaks
├── img/ # README posters: banner + five roadmaps
├── tracks/ # the five tracks, one idea per file
├── tutorials/ # how each lab type is supposed to feel
├── notebooks/ # 20 CPU-first labs
├── reference/ # glossary, books, YouTube, papers, tools
├── assets/diagrams/ # Zorost diagrams (HTML source + PNG)
├── docs/ # assessment of the field, study plans, roadmaps
├── scripts/ # check.py and diagram export
└── .github/ # conduct, security, issues, CI
The field already has a well-known public roadmap tradition. We read it, including Maxime Labonne's widely used llm-course, end to end. We kept the coverage (fundamentals, scientist, engineer) and rewrote every explanation and every picture. We added the modules that 2024 to 2026 made non-optional, we added an operator track and a leader track, and we added labs you can run without a rented GPU.
We do not reproduce third-party course text, notebooks, or artwork. Citations point at the original. The assessment of what was missing is docs/ASSESSMENT.md.
Diagrams use the Zorost house skin (PAPERG paper, INK type, LAVA accent) and the Lab's editorial diagram system. HTML sources sit next to the PNGs so you can change a label without redrawing from memory.
- Read the error. The last line is usually the name of the problem.
- Check that module's knowledge-check answers and the glossary.
- Open a GitHub issue with the file path, what you ran, and the traceback.
- Write info@zorost.com for security or conduct.
See contributing, code of conduct, and security.
Zorost Intelligence designs, ships, and operates AI and data platforms for organizations where accuracy, traceability, and compliance are non-negotiable. This course is developed by the AI Lab and published as Training 02. Training 01 is AI Engineering Lab.
Questions: info@zorost.com.
MIT. See LICENSE. Learn freely, build freely, attribute Zorost Intelligence.
© 2026 Zorost Intelligence LLC · zorost.com · @ZorostAI · info@zorost.com