Practical, code-first AI/ML tutorials for working engineers — version-pinned pipelines, real hardware, and the numbers you need before you run.
🌐 Live site: https://aiml.training
Every tutorial is built end-to-end on a single 12 GB GPU (NVIDIA RTX 3080 Ti), with exact model versions, exact commands, and the real wall-clock times, memory peaks, and eval results from the run — so you know what to expect before you spend the hours.
- Fine-Tune Qwen/Qwen3.5-4B with LoRA on a Single 12 GB GPU — the full loop: quantized base model, 4-bit LoRA training, what fits in VRAM and what doesn't.
- Serving a Fine-Tuned LLM with vLLM (and the LoRA Trap) — get a LoRA adapter serving with vLLM, and the dtype/adapter pitfalls that make it silently fail.
- Evaluating a Fine-Tuned LLM: Perplexity, Win Rate, and a Judge — how to actually measure whether the fine-tune improved anything, not just whether it trained.
- Fine-Tune Kev-4B: A Calibrated Decision Model on a 12 GB GPU — fine-tuning an open decision model (not a chat model) from a published checkpoint, warm-starting from a published LoRA, and running a statistically-significant paired comparison against the stock model.
content/docs/tutorials/— the tutorial pages (one.mdper tutorial)content/docs/_index.md— landing pagelayouts/— Hextra theme overrideshugo.yaml— site config.github/workflows/pages.yaml— GitHub Pages build and deploy
Pre-requisites: Hugo 0.117.0 extended, and Git.
git clone [email protected]:Decent-Software-LLC/aiml.training.git
cd aiml.training
hugo mod tidy
hugo server --logLevel debug --disableFastRender -p 1313The site builds at http://localhost:1313. Drafts (draft: true front matter) only render with hugo server --buildDrafts.
hugo mod get -u
hugo mod tidyThe site deploys to https://aiml.training from the main branch via GitHub Pages. .github/workflows/pages.yaml builds with Hugo 0.117.0 and deploys via the actions/deploy-pages action. In repo settings, the Pages source must be set to GitHub Actions, and the custom domain aiml.training is wired through CNAME.
MIT — the tutorials and site code are MIT-licensed. The model checkpoints and datasets referenced in the tutorials are subject to their own licenses (noted in each tutorial's version-pins section).