Settings

Click Config to open the settings; click it again (or Cancel) to go back to the game.

Presets#

A preset is a set of settings for how the model writes. Pick one in the list next to Config, then open Config to change it.

  • Change a preset: change the values and click Save. This works for every preset, Default included.
  • Make a new preset: change the name at the top of Config and click Save. You can have up to five presets, Default included.
  • Reset to default: the link at the top of Config fills in Roller's built-in values for the preset that's open, whichever it is. Click Save to keep them, or Cancel to leave the preset as it was.
  • Delete a preset: Delete Preset. Default can't be deleted, so there's always one to fall back on.

The settings, in plain words:

  • Temperature: how surprising the writing is. Lower is safer and more predictable; higher is livelier but can wander. 0.7–0.9 is a good range.
  • Top P, Top K, Min P: finer controls for how adventurous word choices are. Best left as they are.
  • Repeat penalty, Presence penalty, Frequency penalty: discourage the model from repeating itself. Small changes only.
  • Max tokens: the longest answer allowed; −1 means no limit. This much room is also kept free in the context (otherwise 1,024 tokens).
  • Seed: leave empty for fresh answers each time; a number makes answers repeatable.
  • Enforce JSON schema: keeps the answers in the form Roller needs. Leave it on, unless a model insists on "thinking" first and fails with it; see The reply was not in the turn format.
  • Stop sequences: words that end an answer. Usually empty.

Port#

Under Port you set the port Roller runs on (default 3211); you open Roller at http://localhost: followed by that number. Change it only if another program already uses it. Roller's helper programs take the next three numbers: llama.cpp (Local mode) the port + 1, Listen the port + 2, and Speak the port + 3, so with 3211 they use 3212, 3213 and 3214.

The change is saved right away and takes effect the next time you start Roller: close its Terminal window and double-click roller.command again. Safari then opens at the new port.

Context size#

The Context size under Context size in Config is how much the models Roller loads can keep in mind; see How long can a game get?. It's used for all Local models, and for LM Studio models that Roller loads itself. A model you loaded in LM Studio by hand keeps the size you gave it there.

A bigger context needs more memory. The change is saved right away, and the model is reloaded with it when you send your next turn.

Roller checks this for you with Local models: after a model has loaded, it compares the context size with what the model and your Mac's memory allow (from llama.cpp's own figures). If the size is far too large (the Mac would start swapping), larger than the model was trained for, or much smaller than it could be, a short note explains it and offers a better size; one click sets it here, and the model is loaded with it at your next turn. Keep keeps your setting, and Roller doesn't ask again for that model and size.

Speech#

Under Speech you choose the voice engine for Speak, the languages to recognise, a voice for each language (▶ plays a sample), and the speed. See Hear lines read aloud.

Nudge rules#

Under Nudge rules you can switch the nudges on and off, change them, add your own with Add rule, or go back to the built-in ones with Restore defaults. Changes are saved right away. What each part of a rule means — the condition, the chance in percent, who gets the impulse — is explained under Nudges.