AI & DataIntermediate
Temperature
PronunciationTEM-prah-chur
Definition
Temperature is a hyperparameter in LLMs that controls the randomness of the generated text. Lower values make the output more deterministic and focused, while higher values make it more creative and unpredictable.
Where you hear it
- Configuring model parameters in an API request. - Fine-tuning or prompting strategies for creative writing tasks. - Debugging unexpected or repetitive model responses.
Examples
Set the temperature to 0.2 for factual tasks to ensure consistency.
Increase the temperature to 0.8 if you want the model to generate more creative stories.
Common mistake
Thinking that a higher temperature means the model is "smarter"; it actually just increases the probability of selecting less likely tokens, which can lead to more hallucinations.