<aside> 💡

Most use cases don't need fine-tuning.

Before you look into fine-tuning, try the following:

1. Better prompts
2. Richer context
3. A more suitable model

Fine-tuning is expensive, brittle, and often unnecessary. It should be the last thing you try, not the first.

</aside>

99% of problems don't require fine-tuning.

Most companies think that fine-tuning is the solution to their problems. In my experience, they could solve almost everything by doing the following:

1. Using a better-suited model for their task
2. Improving their prompts
3. Implementing in-context learning
4. Structuring responses with formatting/schemas
5. Providing examples as part of the prompt
6. Pre-processing and normalizing user inputs
7. Chaining LLM calls with logic or tools
8. Postprocessing outputs
9. Adding guardrails and constraints

Fine-tuning should be your last resort, not the first step.

How to train and fine-tune LLMs yourself?

Fine-tuning is not simple and you should carefully consider other options → great discussion here

Training and Fine-tuning LLMs

Fine Tuning LLMs - Videos

Fine Tuning LLMs - Articles

GPT Services for RAG