Fine-tuning AI models

[widal001]

Topic

Often base models (e.g. ChatGPT 4o, Gemini, etc.) provide great performance out of the box. And the responses from these base models can be improved with Retrieval Augmented Generation (RAG), but fine-tuning a model (adjusting the weights with additional custom data) can provide an additional performance boost and make the model more sensitive to particular use cases.

Questions

• What is the performance difference between:
  • Base models
  • Base models + RAG
  • Fine-tuned models
  • Fine-tuned models + RAG
• How much data is needed to meaningfully tune a model?
• What's the process for tuning a model? How much technical expertise does it require?
• What's the cost associated with tuning a model?

Relevant resources

No response