Fine-Tuning LLMs: From Zero to a Shipped Adapter

Everything you need to actually fine-tune a model and know whether it worked. Covers the decision framework (when fine-tuning beats prompting or RAG), the complete terminology map, LoRA and QLoRA mechanics, dataset construction, a full walkthrough of the self-hosted Unsloth track and the managed Azure AI Foundry track, real published cost figures and a ranked cost-mitigation playbook, evaluation and failure diagnosis, and how to merge, quantise and serve what you trained.

intermediate 150 min 12 lessons

Lessons

What Fine-Tuning Actually Is (And What It Isn't)
The three failure modes behind every bad model output, the four levers you can pull, and the honest decision flow that tells you whether fine-tuning is the right tool.
The Fine-Tuning Family Tree — Every Term You'll Meet
A complete glossary organised into four groups: structural terms (LoRA, QLoRA, adapters), objectives (SFT, DPO, RLHF, RFT), training mechanics (rank, alpha, epochs, packing), and ops terms (JSONL, endpoints, GGUF).
LoRA Under the Hood — Why 7 Billion Weights Fit on Your GPU
The memory maths behind full fine-tuning, what W = W0 + (alpha/r)·B·A actually does, how QLoRA cuts VRAM 4x, and which target modules to adapt.
Data Is the Product
How many examples are actually needed, the four JSONL shapes, the chat-template trap that silently ruins runs, synthetic data via distillation, and a pre-training data checklist.
Track A — A Real Fine-Tune with Unsloth
Unsloth Core vs Studio vs Desktop, exact install commands, and the six-step pipeline: load in 4-bit, attach adapters, load data, apply the chat template, train, merge and export.
Hyperparameters, VRAM and Time — The Numbers
The official VRAM table by model size for QLoRA and LoRA, every recommended hyperparameter with reasoning, the throughput arithmetic for estimating runtime, and where project time actually goes.
Track B — Managed Fine-Tuning on Azure AI Foundry
The supported model and method matrix, the UTF-8-with-BOM data requirement, the upload-create-monitor-deploy workflow in real SDK code, RFT and graders, and the four deployment types.
The Cost Ledger — What It Really Costs and How to Cut It
Real Azure prices for training and hosting, the crossover between a managed endpoint and a rented GPU, a ranked 12-point cost-mitigation playbook, and the hidden costs nobody invoices you for.
Did It Actually Work? — Evaluation and Failure Diagnosis
Baseline-first measurement, the metrics that matter, reading loss curves, a symptom-to-fix diagnostic table, and how to run LLM-as-judge without fooling yourself.
Shipping It — Merge, Quantise, Serve
Three shapes to ship (adapter, merged model, hosted endpoint), how merging works, serving with vLLM and Ollama, multi-LoRA cost savings, and the pre-deployment gate checklist.
Choosing Your Path — Unsloth, Azure, or Something Else
Self-hosted vs managed compared honestly across cost shape, control and ops burden, the hybrid workflow professionals actually use, and the other providers worth knowing.
Grand Quiz — Fine-Tuning LLMs
Twenty-two questions covering all eleven lessons: decision frameworks, LoRA mechanics, datasets, the Unsloth and Azure tracks, cost arithmetic, evaluation, shipping, and strategy.