Michael Limberger

Parameters That Matter

Default settings in most chat UIs are a poor fit for roleplay. Worse, some "fixes" that help general chat models actually hurt Magidonia. This section sets a baseline that matches how the model was trained.

The Big Principle

Magidonia was fine-tuned for creative roleplay. It already handles repetition reasonably well, and it is trained to tell a story with you rather than fight you. So the first rule is: do not fight the model either.

General chat models often loop. People reach for a repeat penalty, DRY (Diverse Random Yielding) samplers, or XTC (eXtendable Thought Complexity) samplers. Those tricks can help bland assistants. On Magidonia they often suppress natural creativity, make common words feel awkward, or nudge the model out of character.

So disable the repetition penalty for Magidonia. Set it to 1.0 (do nothing). Turn off DRY and XTC if your interface offers them. That is a sharp departure from advice that works on other models, and it is the right departure here.

The Recommended Settings

Here is the BeaverAI-tested baseline we use:

Temperature:       1.0
        Top K:             40
        Top P:             0.95
        Min P:             0.02
        Repeat Penalty:    1.0 (disabled)
        Max Tokens:        2048
        Context Window:    16384

Save these. They are the starting point for every Magidonia character in this series. Below is what each dial actually does.

Temperature: The Randomness Dial

Temperature controls how random the next-word choice is. At 0, the model always picks the most likely word: deterministic, often boring, fine for dry technical writing, poor for roleplay. At 1.0 it stays confident but willing to surprise you, which is Magidonia's usual sweet spot. At 1.5 it gets weird in a way that can suit surreal fiction but not grounded scenes.

Start at 1.0. If the character feels too wild, try 0.8. If it feels flat or repetitive, try 1.2. Most of the time you leave it at 1.0.

Top K: Candidate Limiting

Top K limits how many likely next words the model may consider. Top K 10 is narrow and predictable. Top K 40 is wide enough for creativity without losing the plot. Top K 100 is very open and can get chaotic. For Magidonia, 40 is the usual sweet spot. Drop toward 30 if output feels random; try 50 if it feels too predictable.

Top P: Nucleus Sampling

Top P (nucleus sampling) limits by cumulative probability instead of a fixed count. The model adds likely words until it hits the threshold (0.95 means about 95% of the probability mass). When the model is very sure, the pool stays small. When several sensible choices exist, the pool widens. With Magidonia, 0.95 pairs well with Top K 40. Prefer tuning temperature and Top K before you touch Top P.

Min P: The Trash Filter

Min P drops low-probability junk from the long tail of the distribution. At 0.02, words below about 2% probability relative to the top word get filtered, which keeps coherence without smothering creativity. 0.05 is stricter (better for dry technical output). 0.01 is looser (more chaos). For Magidonia roleplay, 0.02 is the baseline. If you see nonsense or invented words, try 0.03 or 0.04. If the model feels boxed in, try 0.01. Adjust in small steps around 0.005.

Repeat Penalty: The Critical Setting

This is the setting that separates Magidonia from most other chat models. A repeat penalty makes the model less likely to reuse a word or phrase. General models often need that help. The Cydonia line was already trained to avoid lazy loops, so stacking another penalty on top makes it dodge ordinary words awkwardly ("the" turns into "said object," "look" turns into "gaze upon").

Keep Repeat Penalty at 1.0 (disabled). Many UIs default to 1.1 or 1.2. Change it to 1.0 and leave it there. If someone tells you to raise it "for better output," they are giving general-model advice, not Magidonia advice.

Max Tokens: Room To Breathe

Max Tokens caps how long a single reply can be. Around 256 you get short dialogue. Around 2048 the model can unfold a scene without feeling rushed, which is our standard for Magidonia. Use 1024 for shorter replies, or 4096 if you want longer passages and can wait for generation.

Context Window: The Memory Limit

Context Window is how much prior conversation the model can see. At 16384 tokens (our recommendation), that is roughly 12,000 to 16,000 words of history, often about 30 to 40 typical turns. That is enough for character consistency without drowning the model in noise.

Very large contexts can actually hurt consistency: at 32K the model may drift off its core traits, and at 4K it forgets too fast. 16384 is the usual goldilocks zone. Short sessions can use 8192. Long campaigns can try 24576. Do not go below 8192 if you care about coherence.

One pattern, several flags

Same baseline, different UIs:

In Ollama CLI (run command):

Show me

ollama run -e TEMPERATURE=1.0 model_name

In OpenWebUI (Settings tab):

Temperature: 1.0
        Top K: 40
        Top P: 0.95
        Min P: 0.02
        Repeat Penalty: 1.0
        Max Tokens: 2048
        Context: 16384

In an API call:

{
          "temperature": 1.0,
          "top_k": 40,
          "top_p": 0.95,
          "min_p": 0.02,
          "repeat_penalty": 1.0,
          "num_predict": 2048,
          "num_ctx": 16384
        }

The Sampler Order

If your interface lets you set sampler order, use this:

1. Min P
        2. Top K
        3. Top P
        4. Temperature

That order filters from strict to loose, then applies randomness to the final pool, which works well for Magidonia.

Tuning Advice

Start with the baseline, then change one thing at a time.

1. Temperature first. It's the most intuitive. Too wild? Drop to 0.8. Too boring? Try 1.2.
        2. Then Min P. Is the model saying weird things? Bump it to 0.03.
        3. Then Top K. Does it feel constrained? Try 50. Too chaotic? Try 30.

Do not touch repeat penalty. Most of the time the defaults above are enough and you can stop tuning.

Quick Reference Table

Parameter	Default	Range	What it does
Temperature	1.0	0.1 to 2.0	Randomness. 1.0 is creative.
Top K	40	10 to 100	How many words to consider
Top P	0.95	0.8 to 0.99	Cumulative probability limit
Min P	0.02	0.01 to 0.05	Filter low-probability junk
Repeat Penalty	1.0	1.0 to 1.2	Keep at 1.0 for Magidonia
Max Tokens	2048	256 to 4096	Response length
Context Window	16384	8192 to 24576	How much history to remember

One More Time

Magidonia is a specialist, not a generic chat model. Do not paste general-purpose sampler advice onto it. The foundation is Repeat Penalty at 1.0. Everything else is fine-tuning around that. Write a character, see how it feels, and only then tweak.