A 3-hour hands-on workshop — Revision 1.4, 10/29/26 Tech Skills Transformations LLC · Brent C. Laster
Pre-trained models are powerful generalists, but most enterprise use cases demand behavior specific to your business: internal terms, company style guides, your own labeling schemes, or regulatory wording. This workshop teaches the complete fine-tuning workflow — preparing the data, training, and checking the results — using models small enough to train in minutes on a CPU-only GitHub Codespace.
Five labs, 10-12 minutes each, all on the same running task: turning free-text support tickets into one strict line in a made-up company's own format.
| Lab | Topic | Model | Method |
|---|---|---|---|
| 1 | Prepare the data, chat templates, score with no training | distilgpt2 (82M) + SmolLM2-360M | none |
| 2 | Full fine-tuning with the Hugging Face Trainer |
distilgpt2 (82M) | all parameters |
| 3 | LoRA with PEFT + the Hugging Face Trainer |
SmolLM2-360M-Instruct | LoRA (trains ~2.3% of the model) |
| 4 | QLoRA (model stored in 4 bits), fold the adapter in and save | SmolLM2-360M + TinyLlama-1.1B | QLoRA |
| 5 | Compare every run, check general skills with standard quizzes, optional comparison with a huge online model | all of the above | evaluation |
- Basic Python proficiency
- A conceptual understanding of what an LLM is
- No prior ML training experience
- No GPU
| Account | Needed? | Why |
|---|---|---|
| GitHub (free) | Yes | To run the workshop's Codespace. Sign up at https://github.com/signup if you do not have one. |
| Groq (free, no credit card) | Optional | Only for Lab 5's optional STEP 6. See below. |
Do this before the workshop starts. It takes about 10 minutes, mostly waiting.
1. Change your Codespace's idle timeout from 30 minutes to something longer.
By default a Codespace stops after 30 minutes without activity, which can happen during a long lecture section. When logged in to GitHub, go to https://github.com/settings/codespaces, scroll to Default idle timeout, set it higher (the maximum is 240 minutes), and click Save.
2. Click the button below to start a new Codespace from this repository.
Click here ➡️ 
3. Click "Create new codespace". (On some screens the button reads Create codespace.)
If VS Code asks whether you trust the authors of the files, click Trust Folder & Continue.
The Codespace then installs the Python libraries and downloads the three models the labs use (about 4 GB), so no lab has to wait on a download later. This takes about 5-8 minutes. It is ready when the terminal shows:
>>> Model prefetch complete.
4. Check that everything works.
In the Codespace TERMINAL, run:
python scripts/verify_setup.py
It takes under a minute. The last line should say
RESULT: ready for the workshop. (Any WARN lines are optional.)
5. Open labs.md.
Open labs.md from the file list on the left. It should open formatted; if
you see raw text instead, press Ctrl+Shift+V (Cmd+Shift+V on Mac). This is the
document you follow during the workshop.
You are ready for the labs!
If your Codespace stops and you need to reopen it: go to https://github.com/codespaces, find the Codespace in the list, click the ... next to it, and choose Open in browser. Your files, the installed libraries, and the downloaded models are all still there.
More detail and fixes for common problems: README-Codespace.md.
Lab 5 ends with an optional step that compares your fine-tuned model against a very large model hosted by Groq. It needs a free Groq API key. Every other lab runs without one, and if you skip this, that step simply prints a note and moves on.
a. In a browser, go to https://console.groq.com and create an account. (If you get an email with a button to confirm, make sure the link opens in the same browser where you were using Groq. If not, copy the link and paste it into the right browser.)
b. In the top right of the Groq screen, click API Keys.
c. Click Create API Key.
d. Give it any name, verify you're human if asked, and click Submit.
e. Copy the key — Groq will not show it to you again.
f. Back in the Codespace TERMINAL, run the command below, paste your key when asked, and press Enter:
source scripts/setup-key.sh
This sets the key for this terminal and for every terminal you open later. (No GitHub settings or Codespace secrets are involved.)
g. Confirm the key works:
python scripts/check_groq.py
You want to see a line starting with OK. If it reports that the model is not
available, Groq has retired it; the script lists what your key can reach and
tells you the one line to change.
You can run the labs on your own computer instead. These guides walk you through
installing everything the Codespace would have set up for you, and explain what
looks different outside the Codespace (including alternatives to the code -d
merge steps):
- Mac: Local_Setup_Mac.md — Apple Silicon Macs (M1 or newer) on macOS 14+. Intel Macs are not supported; please use the Codespace.
- Windows: Local_Setup_Windows.md — Windows 10/11 on Intel or AMD processors. Windows-on-ARM PCs are not supported; please use the Codespace.
Both guides also describe a Docker Desktop option that builds the same environment as the Codespace.
| Codespace | Your own machine (minimum / recommended) | |
|---|---|---|
| CPU | 4-core (chosen for you) | 2 cores / 4 or more |
| Memory | 16 GB | 8 GB / 16 GB |
| Disk | 32 GB | 10 GB free / 20 GB free |
| GPU | none | none — the labs deliberately run on the CPU |
Labs are timed against a 4-core Codespace. On other machines times will differ.
.devcontainer/ Dev container / Codespace configuration
.github/ Copilot instructions for students
common/triage.py Shared helpers: data loading, generation, scoring
data/ The support-ticket dataset (JSONL)
extra/ Completed versions of every lab file (.txt)
finetune/ Lab skeleton files students edit
images/ Screenshots referenced by the docs
scripts/ Setup, model download, environment check, Groq key
labs.md *** THE LAB DOCUMENT — start here ***
Local_Setup_Mac.md Running the labs on your own Mac
Local_Setup_Windows.md Running the labs on your own Windows PC
README-Codespace.md Codespace details and troubleshooting
course-spec.md Course design spec (agenda, topic map, lab plan)
anticipated-qa.md Instructor prep: likely student questions
requirements.txt Python dependencies
| Command | What it does |
|---|---|
python scripts/verify_setup.py |
Checks libraries, downloaded models, 4-bit training, the code command, and the optional Groq key |
python scripts/prefetch_models.py |
Downloads the models (the Codespace does this for you; safe to re-run) |
source scripts/setup-key.sh |
Sets the optional Groq key (Codespace, Mac, Linux, Git Bash) |
.\scripts\setup-key.ps1 |
Sets the optional Groq key (Windows PowerShell) |
python scripts/check_groq.py |
Confirms the Groq key and the Lab 5 model work |
Open labs.md and follow it top to bottom.
Everything the labs save goes in out/ (it is not tracked in git). To start over from scratch:
rm -rf out/
ModuleNotFoundError / wrong Python.
Your terminal prompt should start with (.venv). If not:
source .venv/bin/activate
A model download hangs or fails. The download script is safe to re-run and skips anything already downloaded:
python scripts/prefetch_models.py
If it fails with 429 Too Many Requests, Hugging Face is briefly limiting
downloads (this can happen when a whole class downloads at the same moment).
Wait a few minutes and run it again — no account or token is needed.
Training feels much slower than the lab says.
Check your machine size with nproc --all — it should report 4. (Plain nproc
reports 2 here, because GNU nproc honours the OMP_NUM_THREADS=2 this repo sets
deliberately; see README-Codespace.md.) A genuinely 2-core Codespace will roughly
double every training time in labs.md.
Killed during Lab 4.
Out of memory. Close other running processes and re-run; Lab 4 briefly holds
two models (STEP 6 reloads the 16-bit base to fold the adapter into, and STEP 8 loads
TinyLlama). If it persists, drop QLORA_BATCH from 2 to 1 and skip STEP 8.
Lab 4 fails to load the model in 4-bit.
Run python scripts/verify_setup.py — its 4-bit check tells you whether this
machine can run Lab 4. Codespaces always can. On your own machine, the 4-bit
library needs an Intel/AMD processor with AVX2 (Windows, Linux) or an Apple
Silicon Mac on macOS 14+. If the check fails, use a Codespace for Lab 4 or follow
along with the instructor.
Everything scores 0% after training.
You almost certainly ran the file before merging in the completed code, or you
stopped merging partway. Re-run the code -d step and work top to bottom until
no highlighting is left, save, and run again. The right pane is yours to
edit; the left is the completed reference.
See LICENSE. These materials are for educational use only by attendees of Tech Skills Transformations workshops.
© 2026 Tech Skills Transformations LLC and Brent C. Laster. All rights reserved.
- Web: techskillstransformations.com · getskillsnow.com
- Contact: training@getskillsnow.com