Series — Fine-Tuning, Smallest to Largest (same task, three techniques, smallest model to largest):
Full Fine-Tuning (270M) ← you are here
LoRA (1.5B)
QLoRA (7B)
If the small model worked, why go bigger?
I Fine-Tuned a 270M Model on My Laptop (Full Fine-Tuning, From Scratch)
Part 1 of a 4-part series. Full fine-tuning a tiny Gemma 3 model for intent classification — the generative framing, the loss-masking trick, and why full FT is so learning-rate sensitive.
TL;DRAI
Author fine-tuned Gemma 3 (270M parameters) on Banking77 intent classification, achieving ~96% accuracy via full fine-tuning on a laptop. Full fine-tuning updates all weights (4× memory cost) but is fragile—learning rate 5e-5 stable, 2e-4 crashes—making it the expensive baseline to benchmark LoRA/QLoRA against.
513 words~2 min read
