Sampler Presets

Sampler settings change how the model chooses the next token once it already understands the prompt.

That makes samplers different from wrappers or lorebooks. Wrappers change instructions. Samplers change generation behavior.

Why Samplers Matter

If a model is:

• Too random.
• Too repetitive.
• Too timid.
• Too flat.
• Too verbose.

the sampler is often the next thing to inspect after the prompt.

Built-In Baselines

Abolitus ships with at least two useful baseline styles.

Creative

Designed for looser, more varied output.

Default profile:

• Temperature: 1.0
• Top P: 0.95
• Min P: 0.05
• Top K: 0
• Repetition penalty: 1.1
• Frequency penalty: 0

Good for:

• Expressive roleplay.
• More surprise.
• Wider phrasing variety.

Precise

Designed for tighter, more controlled output.

Default profile:

• Temperature: 0.4
• Top P: 0.8
• Min P: 0.1
• Top K: 40
• Repetition penalty: 1.05
• Frequency penalty: 0.3

Good for:

• Cleaner instruction following.
• More stable structure.
• Less drift.

What the Main Controls Do

Temperature

Higher temperature increases variety and risk. Lower temperature makes output more deterministic.

Top P

Limits generation to the most likely slice of the probability distribution. Lower values usually feel safer and tighter.

Top K

Limits the choice set to the top number of candidates. Useful for making output feel more controlled.

Min P

Filters out very low-probability options. Helpful for reducing weird tail behavior.

Repetition penalty

Discourages the model from repeating itself too aggressively.

Frequency penalty

Pushes the model away from overusing the same tokens across the reply.

Max tokens

Caps how long the completion can become.

Practical Preset Advice

If the model is too flat

Raise temperature slightly before you rewrite your entire prompt.

If the model rambles or drifts

Lower temperature and tighten Top P or Top K.

If the model repeats phrases

Raise repetition penalty a little and consider a small frequency penalty.

If a strong model suddenly feels sloppy

The sampler may be too loose for the scene. High-end models still need sane settings.

Save Presets Instead of Constantly Tweaking Live

If you find a combination that works for a genre, save it as a preset.

Good preset categories include:

• Slow-burn romance.
• Tactical dialogue.
• Narrator-heavy scenes.
• Group-chat banter.
• Small local model stabilization.

Important Caveat

Providers do not always interpret every sampling control in exactly the same way. Treat presets as highly useful cross-route tools, but not as perfect physics constants.