TokenPad

6 tools

LLM Parameter and Reference Tools

Parameters, encodings and behaviour, explained interactively.

Temperature, top-p, max_tokens and stop sequences are set once at the start of a project and rarely revisited, usually to whatever a tutorial suggested.

Each of them has a measurable effect on output quality, latency and cost. These tools show that effect rather than describing it.

Which one you need

The sampling simulator is for understanding what temperature, top-p and top-k actually do to the distribution, which is hard to reason about from the documentation and easy to see once it is drawn.

The max_tokens planner is the practical one. Output is the expensive side and the ceiling has two failure modes: too low silently truncates good answers, unset leaves your bill bounded only by how much the model feels like writing.

Reasoning token estimation matters on any model that thinks before answering, because those tokens are billed at the output rate and never appear in the output. Budgeting from visible answer length can understate the cost several times over.

The general rule underneath this category: the defaults were chosen by someone else for a general audience. Anything that must be reproducible belongs at temperature zero, and that includes every grader.

Questions

What temperature should I use?

Zero for anything that must be reproducible — classification, extraction, routing, and any model acting as a grader. Higher only where variety is the point. The value is also not calibrated across model families, so a setting carried from another provider is a guess.

Should I set both temperature and top-p?

Usually not. They interact in ways that are hard to reason about, and teams that tune both typically end up with a combination they cannot explain or reproduce. Change one and leave the other at its neutral value.

Other categories