A complete, hands-on course that takes you from training a neural net from scratch to LoRA fine-tuning LLMs and image models — running entirely on your own machine.
- Free and open — MIT licensed. No subscriptions, no accounts, no paywalls.
- Runs on almost anything. CPU-only laptop → gaming GPU → workstation. The course detects your hardware and scales itself (see docs/PORTABILITY.md).
- License-clean. Only permissively-licensed material is committed here; models and datasets are downloaded by you at first run, each with its license shown. Full accounting: THIRD_PARTY_LICENSES.md.
- Interactive. Every lesson is a Jupyter notebook with "you should see: ..." checkpoints after each step.
Windows: double-click START HERE.bat → wait for the browser to open.
macOS / Linux: ./start-here.sh
Manual (any OS):
python tools/bootstrap.py --download # sets up env, detects hardware, downloads models
uv run jupyter lab # open the notebooksBootstrap installs uv if needed, creates an isolated environment (your system Python is never touched), picks the right PyTorch build for your hardware, and downloads the course models — printing each item's license as it goes. No internet after first run? Models are cached locally.
No bootstrap? The notebooks also auto-detect hardware on their own. You just need
uv sync (or any Python 3.12 env with the deps in pyproject.toml).
| # | Notebook (under lessons/) |
What you build | Time (GPU) |
|---|---|---|---|
| 1 | 01_foundations/lesson1_tiny_net_from_scratch.ipynb |
First neural net, by hand | ~2 min |
| 2 | 01_foundations/lesson2_tiny_gpt_from_scratch.ipynb |
A tiny GPT that writes fake Shakespeare | ~3 min |
| 3 | 02_finetune_classifier/lesson3_finetune_distilbert.ipynb |
Sentiment classifier at ~95% + overfitting demo | ~5 min |
| 4 | 03_llm_lora/lesson4_lora_qwen3.ipynb |
LoRA-tune a 1.7B chat model into a pirate | ~5 min |
| 5 | 04_diffusion_lora/lesson5_style_lora.ipynb |
Diffusion from scratch + teach image models a public-domain art style | ~15 min |
The through-line: every lesson is the same five lines — data → forward → loss → backward → step. Only the data format, model size, and which parameters move change.
| Your machine | Your experience |
|---|---|
| GPU ≥12 GB VRAM | Everything, including the FLUX.1-schnell bonus track |
| GPU 4–12 GB | Everything except FLUX training (SD1.5 LoRA trains in ~3 GB) |
| CPU only | Lessons 1–3 in full; lesson 4 at smoke-test scale; lesson 5's from-scratch part + pre-computed results |
Details and the "why not Vulkan?" answer: docs/PORTABILITY.md.
- Browser didn't open → copy the
http://localhost:8888/lab?token=...line from the console into your browser. - Out of memory → close GPU-hungry apps (browsers, games) and re-run the cell.
- Download failed → re-run
python tools/bootstrap.py --download(it resumes), or follow the manual fallback instructions indownloads/manifest.json. - License sanity check →
uv run python tools/license_audit.py.
See ROADMAP.md for planned releases (QLoRA, preference tuning, inference & deployment, RAG & agents, more modalities) — shipped incrementally, free as always.
Issues and PRs welcome — see CONTRIBUTING.md (DCO sign-off, no
CLA), CODE_OF_CONDUCT.md, and SECURITY.md.
One hard rule: only permissively-licensed material may be committed (MIT/BSD/
Apache/CC0/public domain) — tools/license_audit.py and tools/check_repo.py enforce
it in CI. When in doubt, make it a runtime download with a manifest entry instead.
More docs: docs/ARCHITECTURE.md (how the repo works), docs/TROUBLESHOOTING.md (help & FAQ), CHANGELOG.md.
MIT — see LICENSE. Third-party components (all downloaded at runtime, never committed) are documented in THIRD_PARTY_LICENSES.md.