This model was finetuned and converted to GGUF format using Unsloth.
Example usage:
For text only LLMs: ./llama.cpp/llama-cli -hf AfkaraLP/gemma-3-4b-coralmc-7000it-v05 --jinja
For multimodal models: ./llama.cpp/llama-mtmd-cli -hf AfkaraLP/gemma-3-4b-coralmc-7000it-v05 --jinja
CoralMC Chat Model Prompting Guide
This model was trained on real chat logs from a multiplayer server.
Its training data uses structured conversation windows and predicts either:
Chat continuation
A response as a specific user
If you prompt it incorrectly, outputs will degrade.
Message Format
Messages in training look like:
[username]: message text
Multiple messages are joined with newlines to form chat history.
Task Type 1: Chat Continuation
System Prompt Structure
[UserA]: message
[UserB]: message
...
[UserN]: message
above is the history of coralmc chat
you will now see a message by NextUser
based on that write how the chat might continue
User Prompt
message from NextUser
Assistant Output
Model outputs several following chat lines:
[UserX]: reply
[UserY]: reply
...
When to use
Use this when you want the model to simulate how the chat continues naturally.
Task Type 2: Respond as Specific User
System Prompt Structure
[UserA]: message
[UserB]: message
...
[UserN]: message
above is the history of coralmc chat
you will now see a message by NextUser
respond to it as TargetUser
User Prompt
message from NextUser
Assistant Output
single reply message written as TargetUser
When to use
Use this when you want the model to answer in-character as a specific username.
Style Expectations
Short informal chat lines
Multiplayer server tone
No narration
No emojis unless chat history already uses them
Usernames preserved in brackets
Example: Continuation
System
[Alex]: anyone online?
[Bea]: yeah what's up
[Alex]: need help building spawn
above is the history of coralmc chat
you will now see a message by Kai
based on that write how the chat might continue
User
I can help if you have materials
Assistant
[Alex]: nice meet at spawn
[Bea]: I’ll grab wood
[Kai]: cool coming now
Example: Respond as Specific User
System
[Alex]: anyone online?
[Bea]: yeah what's up
above is the history of coralmc chat
you will now see a message by Kai
respond to it as Alex
User
server lagging again
Assistant
ugh yeah TPS is dying
Practical Tips
Always include recent chat history
Always include the system instruction text
Keep usernames consistent
Do not add extra roles or narration
Do not remove the bracket format
Intended Use
Server chat simulation
Roleplay continuation
Dataset augmentation
Bot personality testing
Final Note
This model predicts chat patterns.
It is not trained for complex reasoning outside chat simulation.
Available Model files:
gemma-3-4b-it.Q8_0.gguf
gemma-3-4b-it.BF16-mmproj.gguf
⚠️ Ollama Note for Vision Models
Important: Ollama currently does not support separate mmproj files for vision models.
To create an Ollama model from this vision model:
Place the Modelfile in the same directory as the finetuned bf16 merged model
Run: ollama create model_name -f ./Modelfile
(Replace model_name with your desired name)
This will create a unified bf16 model that Ollama can use.
Note
The model's BOS token behavior was adjusted for GGUF compatibility.
This was trained 2x faster with Unsloth
