Fine-Tuning LLMs
For engineers preparing for ML and LLM interviews who need to explain when and how to fine-tune a pretrained model, with a short answer to say and the knowledge behind it for every question. You will be able to answer questions on fine-tuning versus prompting and RAG, LoRA and QLoRA with their memory arithmetic, SFT, DPO, RLHF and GRPO, data quality, forgetting, evaluation, serving and quantisation.
Course Content
Prepares you to explain what fine-tuning is, when to choose it over prompting or RAG, how LoRA and QLoRA work down to the memory arithmetic, and how SFT, RLHF, DPO and GRPO fit together.
Prepares you to answer questions on catastrophic forgetting, safe use of synthetic data, proving a fine-tune really helped, adapter and prompt-tuning methods, and the legal limits of distillation.
Prepares you to answer how fine-tuned models are trained efficiently and served in production: multi-LoRA serving, model merging, QLoRA configuration, loss curves, gradient accumulation, prompt masking, base-model choice and mixed precision.
Prepares you to answer what comes after plain SFT — preference and RL methods, preference data, embedding training with hard negatives — and how to evaluate a fine-tuned model honestly with judges, benchmarks and your own eval set.
Prepares you to answer questions about scale and trade-offs: how much data each kind of training needs, how Mixture-of-Experts models change fine-tuning, how to reduce bias, which fine-tuning APIs to choose, and how NF4 makes 4-bit QLoRA work.