Data
Training example
What is training example?
A training example is one input-and-output pair the model learns from. For supervised fine-tuning it is a prompt and the response you want; the loss is usually computed on the response only, so the prompt teaches context rather than being reproduced.
Consistency across examples matters more than the quality of any one of them. Two examples that give different answers to substantively the same question teach the model that the task is ambiguous, and it will hedge accordingly.
The rule of thumb worth internalising is that the model learns the distribution you show it, including the parts you did not intend. If half your examples end with a question, the fine-tuned model will end half its responses with a question.
Related terms
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.