LLM Fine-Tuning with Unsloth (Done Right)

Fine-tune small open-weight LLMs locally for text tasks. Unsloth + QLoRA makes this practical on consumer GPUs (8 GB VRAM). The workflow: pick a model, format data as instruction-tuning pairs, fine-tune with TRL's SFTTrainer, evaluate against zero-shot baseline, log to MLflow, export to GGUF for llama.cpp deployment.