Skip to content- Sources: primary
- Summary: Code to the Moon fine-tuned Gemma 4 E2B with LoRA in MLX on a JSONL chat dataset to bucket described events by relative day into a fixed JSON schema, with no system prompt needed afterward. The 4-bit fine-tune averaged 0.32 seconds per inference against 14.2 seconds for Gemma 4 31B with thinking, scored 100% on the test set and 98% on a harder unseen set, against about 92% for the 31B model and about 20 to 25% for the base E2B model with instructions. The video warns that fine-tuning is poor at adding new knowledge and can raise hallucination.
- Why it matters: For narrow output-format tasks, a small LoRA fine-tune can replace a large or frontier model and cut latency and token costs by an order of magnitude.