Adapter
Served LoRA
Endpoint plus a report against a frozen baseline.
Model Optimisation & Fine-tuning
Quantisation, distillation, LoRA — and an honest verdict on whether the smaller or tuned model was worth it.
Created by Baljeet Dogra
Course Objectives
- Climb the adaptation ladder only when the rung below is exhausted.
- Run LoRA/QLoRA without overfitting the demo.
- Quantise and measure quality loss, not only size.
- Serve an adapter and write whether it beat prompting on quality, cost and latency together.
What you will produce
Opt
Quantisation card
Size, latency, quality — three numbers, not a blog claim.
6 weeks curriculum
Expand a part for the syllabus. Content stays searchable when closed.
01 The ladder Week 1
Prompt, retrieve, tune. What each fixes. Data quality first.
- Decision framework
- Data
- Splits
- Leakage
02 LoRA / QLoRA Weeks 2–3
A real run, hyperparameters, overfitting, evals including regressions.
- PEFT
- Hyperparams
- Overfit
- Regressions
03 Quantise and distil Weeks 4–5
INT8/INT4, distillation, routing cheap traffic.
- Quantisation
- Distillation
- Routing
- Latency
04 Serve and verdict Week 6
Adapter behind an API. Capstone: worth it or not, in writing.
- Serving
- Cost
- Verdict
- Capstone
Who this is for
Applied GenAI engineers
You can prompt and retrieve. Fine-tuning is still a trophy in your head.
ML engineers
You want PEFT and quantisation with production numbers.
Prerequisites
- Python
- You have called LLM APIs and run a training loop at least once
- A GPU for the LoRA weeks
Not a fit if
- Full pre-training from scratch — that is Custom LLM Training
- People who will not keep a baseline
Related: Applied GenAI Engineering · Custom LLM Training
Questions
Will a “not worth it” verdict fail me?
No. An honest, evidenced no is a pass. Hiding regressions is not.
Which base model?
A small open model you can actually serve. Frontier fine-tunes are discussed, not required.