Перейти к содержимому

Fine-Tuning: When To Train Instead Of Prompt

Jimmy's Tech Deep Dive

0:00 / 0:00

Fine-Tuning: When To Train Instead Of Prompt

29 просмотров · 5 дней назад
Jimmy's Tech Deep Dive
9 подписчиков
29 просмотров · 5 дней назад
Prompt, retrieve, or train? Three different tools - and picking the wrong one costs weeks. A 24-minute breakdown of when fine-tuning actually beats a better prompt. This deep dive explains what fine-tuning really changes inside a model, how it differs from prompt engineering and retrieval-augmented generation, and the four kinds of training people mean when they say the word: supervised fine-tuning, preference tuning, reinforcement fine-tuning and distillation. We look at what a training dataset actually looks like, why consistency beats volume, how LoRA and QLoRA cut the cost enough to run on a single machine, and then work through a four-question decision framework - volume, stability, describability, stakes - applied to real tasks like ticket classification, contract language and strict output formats. It ends with the shape of a sane first fine-tuning project. • Prompting vs RAG vs fine-tuning: which gap each one actually closes • The four flavors of training, including reinforcement tuning and distillation • Catastrophic forgetting, overfitting, and why your eval set comes first • LoRA and QLoRA explained: rank, adapters, four-bit base models • A four-question framework for deciding when to train • When not to fine-tune, and how to run a cheap first experiment Chapters 0:00 Prompt or train? 0:44 What fine-tuning changes 2:06 RAG and the three gaps 4:47 Four kinds of training 6:46 Self-distillation and forgetting 8:19 Building the dataset 10:08 LoRA and QLoRA 13:06 Tooling in 2026 15:07 The decision framework 17:00 When training wins 18:20 When not to fine-tune 19:53 Combining all three layers 21:10 Your first project 22:55 Summary Sources BBVA AI Factory - Prompting or fine-tuning? - https://www.bbvaaifactory.com/prompti... Fine-Tuning, RAG, or Prompt Engineering? LLM Decision Guide - https://moveo.ai/blog/fine-tuning-rag... To fine-tune or not to fine-tune (Meta AI) - https://ai.meta.com/blog/when-to-fine... LoRA vs. QLoRA (Red Hat) - https://www.redhat.com/en/topics/ai/l... MIT's new fine-tuning method lets LLMs learn new skills without losing old ones - https://venturebeat.com/orchestration... Together AI Fine-Tuning: Fine Tune LLM in 12 Steps [2026] - https://tech-insider.org/together-ai-... #FineTuning #LoRA #QLoRA #RAG #LLM