Downloads · 30 days
42
11% of all-time downloads
AfkaraLP/gemma-3-270M-coralmc-1800it-8batch-v11
gemma-3-270M-coralmc-1800it-8batch-v11 is a machine learning model from AfkaraLP. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This model was finetuned and converted to GGUF format using Unsloth.
Downloads · 30 days
42
11% of all-time downloads
All-time downloads
387
Public
Parameters
268M
866 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors536 MB · 62%
From the Hugging Face model README
This model was finetuned and converted to GGUF format using Unsloth.
Example usage:
./llama.cpp/llama-cli -hf AfkaraLP/gemma-3-270M-coralmc-1800it-8batch-v11 --jinja./llama.cpp/llama-mtmd-cli -hf AfkaraLP/gemma-3-270M-coralmc-1800it-8batch-v11 --jinjaThis guide describes how to prompt the model so it stays perfectly aligned with how the training data was constructed. If you follow this, the model behaves. If you don’t, it hallucinates like it’s bored. Choose wisely.
The model expects structured chat transcripts, not freeform prose.
Every message must follow this exact format:
[Username]: Message content
No emojis-as-usernames, no missing brackets, no creative punctuation. The model was not raised in chaos.
All prompts rely on a conversation history block followed by a clearly marked next message.
Context:
[UserA]: message
[UserB]: message
[UserA]: message
### Next Message: [UserC]: message
The model uses this as grounding. Remove it and you get improv night instead of prediction.
The dataset trains two distinct behaviors. Mixing them confuses the model. Humans already do that enough.
Use this when you want the model to continue a chat naturally with multiple speakers.
Context:
<chat history>
### Next Message: [Username]: message
### Continue Chat:
Continue the conversation
[Username]: MessageUse this when you want one specific user to reply.
Context:
<chat history>
### Next Message: [Username]: message
### Next Speaker: [TargetUser]
Generate response
[TargetUser]: message
Exactly one line. If it writes more, it’s showing off.
Because preprocessing already stripped these, do not include them in prompts unless you want degradation:
/login, /tp, !help, . spam)x:123 y:64 z:-456)ok, xd, repeated letters beyond 2)The model was trained in a clean room. Don’t drag mud inside.
The model learned from:
Avoid:
If it sounds like a Discord chat, you’re doing it right.
The model expects:
Dumping 200 lines of history will not make it smarter. It will just get tired.
| Symptom | Cause |
|---|---|
| Model writes narrator text | You didn’t use [Username]: format |
| Model changes usernames | You weren’t consistent |
| Model replies as wrong user | Missing ### Next Speaker |
| Model outputs commands | You reintroduced filtered content |
| Model rambles | Prompt wasn’t constrained |
Context:
[Alice]: anyone on?
[Bob]: yeah
### Next Message: [Alice]: wanna queue?
### Continue Chat:
Context:
[Alice]: anyone on?
[Bob]: yeah
### Next Message: [Alice]: wanna queue?
### Next Speaker: [Bob]
Treat this like a protocol, not a suggestion.
The model is very good at following structure
and very bad at guessing what you meant.
That part is on you.
gemma-3-270m-it.Q8_0.ggufAn Ollama Modelfile is included for easy deployment.
The model's BOS token behavior was adjusted for GGUF compatibility. This was trained 2x faster with Unsloth <img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>