LLM Temperature & Consistency

Replication package studying how temperature settings affect LLM annotation reliability.

A fully reproducible research package testing how OpenAI temperature settings (0.0–2.0) affect consistency and accuracy when using LLMs for rhetorical move-step annotation. Key finding: temperatures of 0.0–0.2 give significantly more consistent results without hurting accuracy; temperatures above 1.6 produce unusable output. Includes all scripts, config, and step-by-step instructions accessible to non-programmers.