Direct questions
Decide whether a specialist model is the right intervention for your task.
- Pages
- 14
- Sources cited
- 15
- Last updated
- 26 October 2026
Should we fine-tune, or improve the system we already have?
Read the answer →START 02Can a small language model replace a frontier model?
Read the answer →START 03RAG versus fine-tuning: which is right for proprietary data?
Read the answer →Keep the answer key outside the experiment.
Examples used to learn the target behaviour. Keep related records together.
Every question, answered.
Should we fine-tune, or improve the system we already have?
Commission adaptation only when a buyer-controlled test shows a valuable gap that a practical API baseline cannot close.
Can a small language model replace a frontier model?
Sometimes, for a bounded task; there is no evidence here for universal replacement.
RAG versus fine-tuning: which is right for proprietary data?
Use retrieval when the immediate need is to supply relevant information at inference time; consider adaptation when repeatable behaviour remains deficient despite adequate context.
How much data do we need to fine-tune an LLM?
There is no universal count that makes a dataset sufficient.
How do we prevent train–test leakage?
Split by the unit that could leak information, not automatically by row.
Which metrics should a buyer insist on?
Define success at the business-task level, with explicit failure constraints.
When is self-hosting cheaper than an API?
Only when its whole operating cost divided by accepted output is lower at the workload and quality level you actually need.
Should we route easy requests instead of fine-tuning everything?
Include routing when a meaningful subset can be handled more cheaply without violating acceptance rules.
What should we own at the end of a supplier engagement?
Specify access and rights asset by asset.
What is the best next action?
Prepare a one-page pilot brief before requesting a production quote.
Frequently asked questions
Short answers to the questions buyers ask.
Read →12 / REFERENCEWhat remains unproven
Evidence gaps and limits that still need resolving.
Read →13 / REFERENCEEditorial self-assessment
The research’s own quality scores and limitations.
Read →14 / REFERENCESources
Trace the section’s claims to the original sources.
Read →Useful evidence.
Visible limits.
Strong fit to a CTO/AI lead decision and broad practical coverage. Evidence supports methods and bounded findings; no private benchmark, measured demand or independent vendor comparison was available.