💭 in your head · nothing to type
You already have a task laid out along the cycle, a climb up the "prompt → context → agents" ladder, and an evaluation plan. One thing is missing — a run of your own.
Picture an evening after which you've shifted a model's weights with your own hands. No paid cloud, on a free GPU in the browser or on your own hardware.
What do you want to see in the "before" and "after" comparison to believe that it worked?
💭 in your head · nothing to type
Keep that answer in mind — it's your success criterion, and it has to exist before the run, not after.
Notice one more thing. Throughout the course you built a sovereign stack: your tools, your context, your agents. A small model fine-tuned on your own data is the next step on the same path.
Consider which of your examples you're willing to hand to someone else's cloud, and which you're not.
This practice is based on Hugging Face's smol-course (Apache-2.0 license). What follows is a condensed retelling with our changes; the original always takes precedence.
One caveat. The current version of the course builds on the larger SmolLM3; some exercises are designed for Hugging Face infrastructure. We need the archived v1 version in the same repository: its exercises use SmolLM2 and have a launch-in-Colab button.
v1/1_instruction_tuning/notebooks/sft_finetuning_example.ipynb — notebook sft_finetuning_example.ipynb: SFT on SmolLM2.v1/3_parameter_efficient_finetuning/notebooks — finetune_sft_peft.ipynb (LoRA) and load_lora_adapter.ipynb (how to load an adapter).v1/4_evaluation — how evaluation works.For deeper theory on transformers and the libraries, see Hugging Face's LLM Course (repository huggingface/course, Apache-2.0).
For those who want to understand how models are trained from scratch, Hugging Face has The Smol Training Playbook. Its license isn't stated, so we only link to it and don't retell it.