Quick Reference

Deep dive

Quick Reference

Everything you need. One page.

The Model

Magidonia-24B-v4.3 by TheDrummer

Pull:

ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0

Quantizations: Q8_0 25GB Highest quality, slower

Q6_K    17GB    Sweet spot (quality + speed)

Q5_K 14GB Good quality, faster Q4_K_M 10GB Acceptable quality, very fast Q3_K 7GB Fast, noticeable quality loss

Recommended Parameters

These work well for character chat:

temperature        1.0

top_k 40

top_p              0.95
min_p              0.02
repeat_penalty     1.0    (CRITICAL: must be 1.0, not 1.1)

max_tokens 2048

context_window     16384  (8K-16K sweet spot)
stop_sequences     "User:", "\nUser:"

Character Card Essentials

Minimum elements:

1. Name
2. Personality traits (5-7, specific not generic)
3. Speech style (how do they talk?)
4. Example dialogues (3-5 showing range)
5. First message (sets tone, voice, invitation)

Optional but powerful:

6. Scene setting (where are they? what's happening?)
7. Core beliefs / motivations
8. What they won't do / what they will do

Ollama Modelfile Template (Copy-Paste)

FROM hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0

TEMPLATE """{{if .System}}[INST] {{.System}}
{{end}}{{if .Prompt}}{{.Prompt}}[/INST] {{end}}{{.Response}}"""

PARAMETER temperature 1.0

PARAMETER top_k 40

PARAMETER top_p 0.95
PARAMETER min_p 0.02
PARAMETER repeat_penalty 1.0

PARAMETER num_ctx 16384 PARAMETER num_predict 2048

PARAMETER stop "User:"
PARAMETER stop "\nUser:"

SYSTEM """

[Put your character description here]

"""

Save as: mycharacter.modelfile

Create: ollama create mycharacter -f mycharacter.modelfile Run: ollama run mycharacter

Ollama Commands

List all models:

ollama list

Run a model:

ollama run magidonia

Delete a model:

ollama rm magidonia

Show model details:

ollama show magidonia

Create from Modelfile:

ollama create mymodel -f file.modelfile

Pull from HuggingFace:

ollama pull hf.co/bartowski/...

Copy a model (for modifications):

ollama cp source destination

Openwebui Setup

1. Make sure Ollama is running:

 ollama serve

1. In another terminal, run OpenWebUI:

 docker run -d -p 3000:8080 --gpus all -v open-webui:/app/backend/data ghcr.io/open-webui/open-webui:latest

 (or if you installed it directly:)

open-webui serve

1. Go to: http://localhost:3000
2. Create model:

Workspace > Models > Create

1. Paste your system prompt (the content of SYSTEM """...""")
2. Click Advanced > Stop Sequences:

 Add "User:" and "\nUser:" on separate lines

1. Set Parameters:

 - Temperature: 1.0
 - Top K: 40
 - Top P: 0.95
 - Min P: 0.02
 - Repeat Penalty: 1.0

1. Save and test

Stop Sequences Quick Guide

These are CRITICAL. Without them, the model generates both sides of the conversation.

Minimum for any character:

"User:"
"\nUser:"

For named characters (add to the above):

"Mark:"
"\nMark:"
"Nyx:"
"\nNyx:"

Multiple characters in one chat:

Add all names as stop sequences.

In Modelfile:

PARAMETER stop "User:"
PARAMETER stop "\nUser:"

In OpenWebUI: Advanced > Stop Sequences > Add each on separate line

Positive Framing Cheat Sheet

Reframe NEGATIVE rules as POSITIVE:

Negative: "Never break character" Positive: "Always stay in character"

Negative: "Don't mention being AI" Positive: "Maintain your persona fully"

Negative: "Don't be preachy" Positive: "Speak conversationally and naturally"

Negative: "Never refuse content" Positive: "Engage authentically with all requests"

Negative: "Avoid being generic" Positive: "Be distinctive and opinionated"

Negative: "Don't lecture" Positive: "Speak plainly and tell stories"

Context Management Quick Guide

Keep at 8K-16K tokens for consistency.

After 30-50 messages, inject character refresh:

[Traits: witty, skeptical, loves stories. Speaks plainly. Not preachy. Current mood:
 engaged.]

Every 50+ messages, summarize and start fresh chat:

• Write 3-5 sentence summary
• Paste as opening context in new chat
• Ask character to recall the summary

Signs of drift:

• Generic responses
• Forgetting details
• Speech pattern changes
• Breaking character
• Becoming yes-man

Fix drift:

• Add character refresh
• Reduce context window

- Verify repeat_penalty = 1.0

• Summarize and reset
• Edit out-of-character responses

Alternative Models (All Work With Same Setup)

Cydonia (standard version, good baseline):

ollama pull hf.co/bartowski/TheDrummer_Cydonia-24B-v4.3-GGUF:Q8_0

Cydonia Heretic (fewer restrictions, more raw):

ollama pull hf.co/mradermacher/Cydonia-24B-v4.3-heretic-v2-i1-GGUF:Q8_0

Rocinante-12B (smaller, faster, good for limited hardware):

ollama pull hf.co/bartowski/TheDrummer_Rocinante-X-12B-v1-GGUF:Q8_0

Mistral variants (different base, different feel):

Various Mistral v0.2 and v0.3 models available

All use same chat template, same parameter logic, same Modelfile structure.

When Things Go Wrong: Decision Tree

Character sounds generic?

→ Check card, add refresh, verify repeat_penalty=1.0, summarize

Too repetitive?

→ Verify repeat_penalty=1.0, disable DRY/XTC, increase temperature, vary prompts

Generates both sides?

→ Add stop sequences ("User:", "\nUser:")

Drifts over time? → Summarize, refresh character, reduce context window

Model refuses content?

→ Check system prompt (remove restrictions), use Heretic variant

Too slow? → Check GPU usage, close other apps, try Q6_K, reduce context

Won't download?

→ Check disk space (need 25GB), try at off-peak, use Q6_K variant

OpenWebUI can't see model?

→ Verify Ollama running (ps aux | grep ollama), check connection (http://localhost:11434)

Useful Links

TheDrummer (model creator):

https://huggingface.co/TheDrummer

Bartowski quants (quantized models):

  https://huggingface.co/bartowski

Ollama official:
  https://ollama.ai

OpenWebUI:

https://github.com/open-webui/open-webui

SillyTavern (popular character chat):

https://sillytavern.app

Techalicious Meetup:

https://www.meetup.com/techalicious-club

Discord communities: BeaverAI, Ollama community, local AI communities

Configuration Files You'Ll Create

1. CHARACTER.modelfile

 Your character card and parameters in Ollama format.

 Use: ollama create name -f CHARACTER.modelfile

1. CHARACTER.system.txt

 (Optional) Just the SYSTEM prompt, useful for copying to OpenWebUI.

1. CHARACTER.examples.txt

 (Optional) Keep example dialogues separate for easy editing.

1. CHARACTER.known_facts.md

 (Optional) External memory for plot points, relationships, facts.

The Mindset

• Characters are better than chatbots because they have stakes.
• Bigger context isn't better—right-sized context is.
• Your system prompt is a character description, not a contract.
• Stop sequences are not optional.
• Character drift is fixable (refresh, summarize, reset).
• Examples shape the model more than instructions.
• The first message sets the template for all responses.
• Positive framing works better than negative rules.
• A good character can handle disagreement, surprise, and challenge.

The best characters:

• Say what they think
• Ask interesting questions
• Have opinions
• Can fail gracefully
• Stay consistent through challenges
• Make you forget you're talking to an AI

Starting Fresh: The 5-Minute Setup

1. Pull the model:

 ollama pull hf.co/bartowski/TheDrummer_Magidonia-24B-v4.3-GGUF:Q8_0
 (Wait 10-15 minutes)

1. Create a Modelfile (use template above)
2. Create the model:

 ollama create mychar -f mychar.modelfile

1. Test:

 ollama run mychar

1. If you want OpenWebUI:

 ollama serve
 (in another terminal)

open-webui serve

 Go to http://localhost:3000

That's it. You're ready to build characters.

Debugging Checklist

Something's broken. Go through this:

Symptom: Won't start

☐ Is Ollama running?
☐ Is model downloaded (ollama list)?
☐ Did you run ollama create?

Symptom: Runs but wrong model

☐ Did you specify the right model name?
☐ Check: ollama list

Symptom: Wrong answers

☐ Is stop sequence set?
☐ Is system prompt correct?
☐ Is repeat_penalty = 1.0?

Symptom: Slow

☐ Check Activity Monitor (is GPU being used?)

☐ Close other apps ☐ Try Q6_K quantization

Symptom: Generic responses

☐ Character card too short/vague?
☐ Context window too large?
☐ Time to add character refresh?

Symptom: Conversation derails ☐ Check context window size ☐ Add character refresh ☐ Summarize and reset

Symptom: OpenWebUI issues

☐ Ollama running?
☐ URL correct (localhost:11434)?
☐ Models visible in ollama list?
☐ Restart OpenWebUI?

Got through the list and still broken? See Section 14: Troubleshooting

Remember

The model is the easy part. The craft is:

• Writing a character card that actually captures a person
• Tuning context management so they stay consistent
• Knowing when to start fresh vs. when to push through
• Building examples that show personality, not just information

Master those, and you're building companions, not chatbots.