Temperature adjusts how the model selects among probable next tokens. Low temperature makes the model favor the most likely continuation, producing consistent output. Higher temperature increases variability and creativity.
Values near zero produce near-deterministic responses for the same input.
Alternative Names:
Temperature, Sampling Temperature
Why it Matters?
Legal applications generally want low temperature, because consistency matters more than variety when extracting facts or summarizing testimony. Two runs producing different summaries of the same deposition is a quality problem, not a feature. Where a vendor's output varies noticeably across identical requests, temperature configuration is worth asking about, since it affects whether validation results are reproducible.
Frequently Confused with
Related terms
Frequently asked questions
What temperature suits legal applications?
Why does temperature affect validation?
