Fine-Tuning LLMs
SFT, DPO and LoRA on open weights, with evals to prove it helped.
Fine-Tuning LLMs is a self-paced online course: 9.5 hours of recorded sessions across 6 modules, with 5 graded by Nova, Square 1's AI tutor. The founding price is A$45, down from A$90. Start any time and keep your own pace.
- Module 1When to fine-tune60 min
- Module 2Data for post-training90 min · graded
- Module 3Supervised fine-tuning with LoRA120 min · graded
- Module 4Preference tuning90 min · graded
- Module 5Serving and cost90 min · graded
- Module 6Project: your model, deployed120 min · graded
- recorded hours
- 9.5
- modules, in order
- 6
- deployed projects
- 4
- graded checkpoints
- 5
Who it is for, and what you receive.
Developers with Python and some PyTorch who want to fine-tune open weights and prove it helped. A GPU budget is stated up front.
Python and some PyTorch; a GPU budget we specify.
- 1
The recorded sessions
Taught by an instructor, taken in order at your own pace, yours for twelve months.
- 2
Graded work after every module
Exercises and projects marked by Nova against a rubric you can read, with what you did well and what to fix.
- 3
The projects
Deployed and graded, each one on your record with its repository.
- 4
Your hiring plan
The same six-part plan the bootcamps use, run with the career agent.
- 5
The record and the certificate
Every grade at square1ai.com/u/{handle}; a credential ID that resolves at square1ai.com/verify.
- 6
Nova as your tutor
Help that is about your actual work, because it has read all of it.
The content plan, module by module.
9.5 recorded hours from an instructor, taken in order at your own pace. Nova grades the work at the end of each module.
-
Prompting versus fine-tuning
Task types
Budgets
The projects.
Each is deployed and graded by Nova against a rubric you can read before you start.
- Project 1
A training set
You are preparing to fine-tune an open-weights model to answer customer questions for a fictional home-insurance broker. Curate a training set from public and synthetic examples, recording where each came from and under what licence, and removing duplicates and anything you cannot trace.
You hand in- Curated training set
- Provenance and licence sheet
- Short curation write-up
The rubric requiresEvery example has a source and a licence.
- Project 2
An SFT model
Using your curated set, fine-tune an open-weights model with LoRA for the broker's question-answering task. Compare it against the untuned base model and a well-prompted baseline, running evals during training so you can see when it stops improving.
You hand in- LoRA adapter and training code
- Held-out eval report
- Training curves with notes
The rubric requiresThe fine-tuned model beats the baseline, with the gain measured on a held-out set.
- Project 3
A preference-tuned model
Build a preference set of chosen and rejected answers for the broker's questions, then run DPO on your SFT model. The hard part is a fair win-rate eval: blind comparisons against the SFT model on prompts it never trained on.
You hand in- Preference set with provenance
- DPO run and config
- Win-rate eval report
The rubric requiresWin rate against the SFT model is measured on held-out prompts and reported.
- Project 4
A deployed model
Quantise your best model, serve it behind an endpoint and put monitoring on quality and latency. Then cost it honestly against a hosted API doing the same job at the broker's expected volume, including the hours you spend running it.
You hand in- Deployed model endpoint
- Monitoring dashboard
- Cost comparison against hosted API
- Short final report
The rubric requiresThe model is deployed with monitoring and a cost report comparing it to a hosted API.
What you can do at the end.
A post-training method from data to serving, and four graded projects on your record.
- 1
Decide when fine-tuning beats prompting.
- 2
Build a training set with provenance.
- 3
Fine-tune with LoRA and beat a baseline.
- 4
Preference-tune with DPO and measure win rate.
- 5
Serve the model and compare its cost to a hosted API.
No placement rate is shown, because there are no graduates to count yet. The roles above are what the projects are built for.
Your record at the end.
Every exercise and project is graded by Nova against a rubric you can read, and every grade is kept on one page an employer can open and run. This is what the programme writes to it.
- Graded, line by line
- Nova reads every submission against the brief and the rubric and returns a score, what you did well and what to fix.
- Module by module
- Each module ends in graded work; the next opens when you are ready, on your own schedule.
- Nova remembers
- Help is about your actual work, because the tutor has every submission and every failed exercise of yours.
- One page an employer can run
- Every grade and project at /verify. An employer opens it and runs the code.
- ProjectGraded
A training set
Every example has a source and a licence.
- ProjectGraded
An SFT model
The fine-tuned model beats the baseline, with the gain measured on a held-out set.
- ProjectGraded
A preference-tuned model
Win rate against the SFT model is measured on held-out prompts and reported.
- ProjectGraded
A deployed model
The model is deployed with monitoring and a cost report comparing it to a hosted API.
The entries, not the grades: those are yours to earn. The page lives at /verify and an employer needs no account to open it.
6 entries, at your own pace. Reserving a spot holds the founding price.
Pre-orderHow we help you find a job.
Proof, not a certificate: the projects you deployed are the thing you show, and the tools below are yours to use.
The whole plan on a few pages: the parts, the projects, the gates and the price. We email it to you.
Get the booklet- 1
Your hiring plan
The same six-part plan the bootcamps use: target roles, the gap map, the proof to send, weekly actions, an interview log, the outcome. You run it with the career agent.
- 2
A record an employer can run
Your graded projects on /verify. An employer opens it and runs the code.
- 3
The career agent
Paste a real job posting at /career and it maps the role to your graded work and what to do next.
- 4
The roles directory
Every role we prepare people for, what it pays and what it asks, at /roles.
- 5
A path to the live cohort
If you want the instructor, the gates and the hiring sprint, the bootcamp on the same subject is one waitlist away.
Questions people ask.
About Fine-Tuning LLMs, answered from the plan on this page.
- Fine-Tuning and Open-Weights Bootcamp · 12-week live bootcamp
- Forward Deployed Engineer Bootcamp · 12-week live bootcamp
- LLM Evals and Reliability Bootcamp · 12-week live bootcamp
Founding spots are open. Hold yours.
Recorded by an instructor who does this work, graded by Nova. Spot holders hear the opening date first, and nothing is charged before you confirm.
- A post-training method from data to serving
- Four graded projects on your record