When is fine-tuning a foundation model the wrong first step?
Why B is the answer
When the model only lacks access to current company documents
Missing facts are a retrieval problem. Tuning changes behavior and does not give the model new knowledge.
In detail
- Missing documents are a retrieval problem.
- Fine-tuning does not add current facts.
- Tuning changes behavior, not knowledge.
Use case
An illustrative scenario, not a real incident.
An assistant lacks this quarter's pricing. Tuning would not add it, while retrieval would.
See it in different ways
stateDiagram-v2 [*] --> Prompting Prompting --> Rag: lacks current facts Prompting --> FineTuning: quality or format is not reached Prompting --> Done: meets the need Rag --> Done FineTuning --> Done
flowchart TD
Q{"Why is the answer wrong?"}
Q -->|"Facts"| A0["Add retrieval"]
Q -->|"Style or format"| A1["Consider tuning"]Key point: Missing facts are a retrieval problem.
Why the other options are wrong
- A. When prompting alone has been tried and has stopped improving results
- That is a good reason to fine-tune.
- C. When output needs a very specific style that prompts cannot hold
- That is a good reason to fine-tune.
- D. When a task needs consistent domain behavior across many requests
- That is a good reason to fine-tune.