Skip to main content

Preset Config Menu

❓What Are Model Presets?​

Model Presets are predefined configurations that bundle commonly used settings for AI models.

⭐ 3 Things You Need to Know​

  1. Create a Preset for each model and task
    • Every Model needs different settings → PLEASE MAKE SURE TO carefully read the Performance - GPU Section
    • Every Task needs different settings
  2. Is your task creative, or do you need a more focused response?
    • Focus on the Temperature and TopP setting (e.g the deterministic (focused) preset has low Temperature and TopP).
    • If you want to learn more about it, learn more about Samplers.
  3. (Max) Context Size sets the max amount of tokens your model gets fed with
    • This value can be too short for e.g. a very large Char. Make sure to understand what is a Context - It is one of the most important concepts in Language Models that you want to understand.
    • Increase it according to what the model supports (see Model Menu), but be aware, it affects token generation speed significantly.

Performance - GPU​

Use either GUESS SETTINGS or the Wand Button if you use a GPU runtime (see Settings). If you already have experience, you can set the number of layers manually.

❓Why So Many Settings?​

AI models are highly flexible tools, and small changes in their configuration can dramatically affect how they respond. These settings exist to let you fine-tune the model for:

  • Tone and style (e.g. creative vs. factual)
  • Performance (e.g. short replies vs. long reasoning)
  • System resources (e.g. reducing context size on limited hardware)

While the number of options may seem overwhelming at first, presets and defaults help make the process approachable — and they give power users the control they need to get the best results.

Prompt Format​

If you change the Text AI Model, you also need to set the correct prompt format. Why? Because there is no universal standard for how models are fine-tuned to behave in a chat-like manner — each model builder handles it differently.

Setting the appropriate prompt format helps ensure that the model understands and follows a conversational structure, which is especially important for smaller or less capable models.

tip

If you're unsure, use the "Model Recommended" format. If that’s not available, either leave it blank (no format) or experiment to see what works best.

❓What Are Samplers?​

Samplers define how the model selects the next word or token when generating a response. They influence the creativity, diversity, and predictability of the output.

For example, if you're looking for a factual summary or trying to improve code, it's best to avoid samplers designed to reduce repetition or prioritize novelty—since these can introduce unnecessary variation. However, those same samplers are valuable for tasks like roleplay or storytelling, where variety and unpredictability enhance the experience.

Important Samplers​

Standard Samplers​

  • Temperature: Think of this as a randomness dial. Higher values (>=1.0) let the AI be more adventurous with its word choices. Lower values (0.2) make it stick to safer, predictable words. At Temperature = 0, it always chooses exactly the most likely next word (purely greedy).
  • Top-p: Imagine sorting all possible next words from most to least likely and then picking just enough words so their combined chances reach a certain percentage p (like 90%). The AI then picks from that smaller set.
  • Min-p: It throws out any words that are below a percentage of the top choice’s chance. So if the best guess has 50% probability and your min‑p is 0.1, words with less than 5% chance get removed. Keeps things on track without strangling diversity.
tip

This site provides a good Visualization🔗

Additional Samplers​

  • Exclude Top Choices (XTC): Sampler for creative tasks, occasionally (based on the probability setting) removes the highest‑probability tokens (top choices), then samples from the remainder to promote creativity and avoid clichés.
  • Don't Repeat Yourself (DRY): Detects if continuing with a given token would extend a previously seen sequence in the prompt or generation history and applies an exponentially increasing penalty to discourage repetition.
SettingWhat It Does for You
TemperatureControls how adventurous the text feels
Top‑pKeeps a dynamic pool of plausible next words
Min‑pFilters out unlikely words based on top‑choice strength
Exclude Top Choices (XTC)Occasionally removes top picks to spark creativity
Don’t Repeat Yourself (DRY)Stops the AI from parroting phrases it already used

There are more than these, but these are all you need.

info

Different samplers suit different tasks — creative writing benefits from more randomness, while technical tasks like summarizations, translations, transformations, ... may prefer more predictable sampling. You can use no sampler, a combination of samplers, or all of them.