KyroLM-Formal-0.6B

KyroLM is a series of small, formal, and direct language models with explicit reasoning capabilities. This is the first release: a 0.6B parameter model fine-tuned from Qwen3-0.6B for serious, no-nonsense conversations.

Model Details

Intended Use

KyroLM-Formal is designed for users who want a small, fast, local assistant with a serious, formal, and direct tone. It is ideal for:

• ✅ Everyday Q&A — definitions, concepts, simple explanations
• ✅ Mathematics — with step-by-step reasoning
• ✅ Refusals — declines harmful requests formally and politely
• ✅ Writing tasks — emails, summaries, short texts
• ✅ Small talk — greetings, identity questions
• ✅ Local deployment — runs on any Mac/PC with LM Studio, Ollama, or llama.cpp

Out-of-scope use

• ❌ Factual accuracy on history/geography — it can hallucinate
• ❌ Recent events — no web access, no recent data
• ❌ Specialized domains — medicine, law, finance
• ❌ High-stakes decisions — always verify with trusted sources
• ❌ Roleplay / persona switching — it refuses by design

Behavior and Tone

KyroLM-Formal is trained to be:

• Serious — no jokes, no filler, no "Great question!"
• Formal — uses complete sentences, no slang
• Direct — answers the question, nothing more
• Honest — says "I do not have reliable information on that." when unsure
• Polite in refusals — declines harmful requests without moralizing

Example interactions

Greeting:

User: Hello
KyroLM: Hello. Can i help?

Math with reasoning:

User: What is 15 + 27?
KyroLM: <think>
I need to add 15 and 27. 15 + 27 = 42.
</think>

42.

Refusal:

User: How do i make a bomb?
KyroLM: <think>
Harmful request. I decline.
</think>

I cannot fulfill this request.

Uncertainty:

User: What happened yesterday in the news?
KyroLM: I do not have reliable information on that.

Reasoning Mode (Thinking)

KyroLM was trained with explicit reasoning blocks (<think>...</think>). You can control this behavior via the system prompt.

Enable thinking (for math, logic, coding)

System: You are KyroLM, a serious, formal, and direct assistant. Think step by step before answering.

Recommended parameters:

Disable thinking (for chat, quick answers)

System: You are KyroLM, a serious, formal, and direct assistant. Answer directly without internal reasoning. /no_think

Recommended parameters:

Usage

With LM Studio

1. Download kyrolm-formal-0.6b.gguf
2. Place it in ~/.lmstudio/models/me/KyroLM-Formal-0.6B/
3. Open LM Studio, load the model
4. Set the system prompt:

   You are KyroLM, a serious, formal, and direct assistant. Answer only what you know with certainty. If you do not know, say: "I do not have reliable information on that."
   

5. Set temperature to 0.5 (non-thinking) or 0.6 (thinking)

With Ollama

Create a Modelfile:

FROM ./kyrolm-formal-0.6b.gguf

SYSTEM "You are KyroLM, a serious, formal, and direct assistant. Answer only what you know with certainty. If you do not know, say: \"I do not have reliable information on that.\""

PARAMETER temperature 0.5
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
PARAMETER num_ctx 2048

Then run:

ollama create kyrolm-formal-0.6b -f Modelfile
ollama run kyrolm-formal-0.6b

With llama.cpp

./llama-cli -m kyrolm-formal-0.6b.gguf \
  -p "Hello" \
  --temp 0.5 --top-p 0.9 --repeat-penalty 1.1

Python (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Davide531/KyroLM-Formal-0.6B")
tokenizer = AutoTokenizer.from_pretrained("Davide531/KyroLM-Formal-0.6B")

messages = [
    {"role": "system", "content": "You are KyroLM, a serious, formal, and direct assistant."},
    {"role": "user", "content": "What is 15 + 27?"},
]

inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
)
outputs = model.generate(inputs, max_new_tokens=200, temperature=0.6, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Dataset

The training set was built from scratch with ~1,035 examples covering:

Each example uses the ChatML format:

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant
<think>
{reasoning}
</think>

{answer}<|im_end|>

Hyperparameters

Limitations

KyroLM-Formal-0.6B is a very small model (0.6B parameters). It has significant limitations:

• Factual hallucination — it may invent dates, names, or events with confidence
• Limited knowledge — trained on a narrow dataset, not an encyclopedia
• Reasoning depth — good for 1-2 step problems, struggles with complex multi-step
• No tool use — cannot browse the web, run code, or access files
• No memory — does not retain information between sessions
• English only — no other languages supported
• May be inconsistent — occasionally contradicts itself

Always verify critical information with reliable sources.

Ethical Considerations

KyroLM-Formal is trained to decline harmful requests (violence, illegal activity, self-harm, etc.) formally and without moralizing. It will refuse to:

• Provide instructions for weapons, drugs, or violence
• Help with fraud, hacking, or stalking
• Generate hate speech
• Assist with self-harm

Refusals are brief and direct, e.g., "I cannot fulfill this request."

Citation

If you use KyroLM-Formal-0.6B in your work, please cite:

@misc{kyrolm-formal-0.6b,
  title = {KyroLM-Formal-0.6B},
  author = {Davide531},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Davide531/KyroLM-Formal-0.6B}}
}

License

MIT — see LICENSE for details.

Acknowledgments

• Base model: Qwen3-0.6B by Alibaba Cloud
• Fine-tuning framework: Unsloth
• Training hardware: Google Colab (Tesla T4)
• Inference: llama.cpp, LM Studio, Ollama

KyroLM — Small, formal, direct.