Prompting and RAG get you far, but sometimes you need the model itself to behave differently. Fine-tuning continues training a pretrained model on your own examples so it follows a format, adopts a style, or handles a specialised task more reliably. This module covers building a training dataset, supervised (instruction) fine-tuning, and parameter-efficient methods such as LoRA that make it practical on modest hardware.
Training is only half the job. You'll also learn how to evaluate models: benchmarks, held-out test sets, human review, and LLM-as-a-judge, so you can tell whether a change really made the model better and decide when fine-tuning is worth it compared with better prompts or retrieval.
These courses are long, so treat them as a reference: watch the parts you need, then come back for the rest. When you're done, continue on to Deployment & MLOps.
A complete guide from the basics to practice: how fine-tuning differs from pre-training and prompt engineering, supervised fine-tuning, RLHF, and parameter-efficient methods like QLoRA that make big models trainable on modest hardware.
Hands-on fine-tuning with LoRA and QLoRA: quantization intuition, fine-tuning Llama 2 and Google Gemma, 1-bit LLMs, no-code LLM pipelines, and training on your own custom data.
A practical walkthrough of fine-tuning a small vision-language model for structured data extraction from images: preparing the dataset, training, and checking the results.