Skip to content

Presets & sampling

A preset bundles the whole "engine tuning" of a chat — its system-prompt template, its post-history block, optional impersonate/continue overrides, and optional per-preset sampling — so you can swap the entire feel of a roleplay in one tap.

What a preset is

A preset is a named package of:

  • A system-prompt template — the instructions that frame the roleplay, written with template tokens that fill in the active character, persona, and lore.
  • A post-history block — instructions placed after the conversation history (the chara_card post_history_instructions slot), for rules you want kept close to the model's next reply.
  • Optional impersonate and continue overrides — alternate prompts used when Pyre writes your turn for you or extends a reply.
  • Optional per-preset sampling overrides — tuning numbers that apply only while this preset is active.

Switch presets and the prompt scaffolding and tuning change together.

Template tokens

The system-prompt template (and the other prompt slots) understand template tokens that Pyre resolves at send time — for example {{char}}, {{user}}, {{description}}, {{scenario}}, and {{wiBefore}} / {{wiAfter}} for fired lorebook entries.

The full list of tokens and what each resolves to lives in the Macros & template tokens reference.

Per-preset sampling overrides

A preset can pin its own sampling. Each override is optional, and it overrides your global Model Settings only when set — leave one blank and the global default applies.

Override What it controls
Temperature Randomness / creativity of the output.
Top-p Nucleus sampling cutoff.
Top-k Limits sampling to the top k candidate tokens.
Max tokens Upper bound on the reply length.
Frequency penalty Discourages repeating the same tokens.
Presence penalty Discourages reusing tokens that already appeared.
Min-p Minimum probability threshold for candidate tokens.
Top-a Adaptive cutoff based on the top token's probability.
Repetition penalty Penalizes repeated tokens to reduce loops.

Tip

Because each number is optional, a preset can tune just one or two knobs (say, a lower temperature) and inherit everything else from your global settings.

The locked "Pyre Default" preset

Pyre ships with a tuned, sealed preset — "Pyre Default" — a sane roleplay Gamemaster starting point with third-person/present-tense formatting rules, a "the character is law" instruction, and modern sampling defaults. It can't be edited or deleted, so there's always a known-good baseline to fall back to.

To customize it, copy it to fork: the copy is fully editable, and you can tune its prompt and sampling however you like.

Global Model Settings

Underneath every preset sits your global Model Settings — temperature, top-p, top-k, max tokens, and the streaming toggle. These are the baseline every chat uses unless a preset's overrides replace them.

Note

The old short-term "memory" message-trim slider has been removed. Long-Term Memory is now Pyre's context manager — see Long-term memory for how the running recap keeps long chats coherent.

See also