Decoding Parameters and Output Control

Core controls

Parameter	What it controls	Typical usage
temperature	randomness of sampling	0.2-0.4 factual, 0.8-1.2 creative
top_k	sample from top-k candidates	conservative diversity
top_p	sample from nucleus probability mass	adaptive diversity
repetition/frequency penalty	reduce repeated tokens	long outputs, summaries
max_tokens	hard output length cap	cost and latency control
stop sequences	forced termination string	structured response boundaries

Practical presets

• Factual Q&A: low temperature + low top_p + moderate max_tokens.
• Creative writing: higher temperature + top_p around 0.9-0.95.
• Code generation: low temperature + stop sequences for fences/turn boundaries.
• Brainstorming: higher temperature + larger token budget.

Important tuning note

Avoid extreme values on both top_k and top_p at the same time. Tune one primary sampling control first, then temperature.