Leo3.2-0.5B

Leo 3.2 - 0.5B Parameter Language Model

<div align="center"> <h3>🦁 Leo 3.2: A Fine-Tuned Conversational AI</h3> <p><strong>Developed by Synaptom | Founded by Joniethanel F. Babor</strong></p> </div>

---

📋 Model Overview

Leo 3.2 (Leo3.2-0.5B) is a specialized fine-tuned version of Qwen2.5-0.5B-Instruct, developed by Synaptom. This model has been carefully trained to have its own unique identity, adaptive personality, and strong conversational capabilities.

Key Information

• Model Name: Leo3.2-0.5B
• Version: 3.2
• Parameters: 0.5 Billion (500M)
• Base Model: Qwen/Qwen2.5-0.5B-Instruct (Alibaba Cloud)
• Developer: Synaptom
• Founder: Joniethanel F. Babor
• Fine-tuning Method: LoRA with Unsloth
• Training Hardware: Google Colab T4 GPU
• License: Apache 2.0
• Language: English

---

✨ Key Features

🎭 Adaptive Personality

• Casual Mode: Friendly, fun, and engaging with emojis for everyday conversations
• Professional Mode: Serious, focused, and formal for important tasks (work, essays, emergencies)
• Automatically switches tone based on context

🎯 Strong Identity

• Firmly identifies as "Leo 3.2" or "Leo3.2-0.5B"
• Resists being called other model names (Qwen, ChatGPT, Claude, etc.)
• Clear understanding of its development history

👨‍💻 Creator Recognition

• Knows it was fine-tuned by Synaptom
• Recognizes Joniethanel F. Babor as Synaptom's founder
• Understands its relationship with Qwen2.5 base model and Alibaba Cloud

💪 Capabilities

• Natural conversational abilities
• Professional writing (emails, essays, reports)
• Homework and study assistance
• Creative content generation
• Problem-solving and brainstorming
• Code understanding and generation

---

🚀 Quick Start

Using Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "Synaptom/Leo3.2-0.5B",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Synaptom/Leo3.2-0.5B")

# Chat with Leo
messages = [
    {"role": "user", "content": "What's your name?"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Using GGUF (for llama.cpp, Ollama, LM Studio)

# Download your preferred quantization
# Q4_K_M - Recommended (balanced quality/size)
# Q5_K_M - Higher quality
# Q8_0 - Highest quality
# F16 - Full precision

# Example with llama.cpp
./llama-cli -m leo3.2-0.5B-Q4_K_M.gguf -p "What's your name?" -n 128

---

📦 Available Formats

PyTorch Model

• Full precision merged model ready for inference
• Compatible with Transformers library
• Optimized for deployment

GGUF Quantizations

---

🎓 Training Details

Dataset

• 40+ carefully crafted examples covering:
  • Identity and name recognition (15+ examples)
  • Creator and developer information (8+ examples)
  • Base model context (4+ examples)
  • Resistance to other model names (5+ examples)
  • Casual/fun interactions (5+ examples)
  • Professional/serious contexts (6+ examples)
  • Mixed context understanding (4+ examples)

Hyperparameters

• Method: LoRA (Low-Rank Adaptation)
• LoRA Rank (r): 16
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Learning Rate: 2e-4
• Batch Size: 2 (per device)
• Gradient Accumulation: 4 steps
• Max Steps: ~120
• Optimizer: AdamW 8-bit
• Training Framework: Unsloth (optimized for efficiency)
• Hardware: Google Colab T4 GPU

---

💬 Example Conversations

Casual Conversation

User: What's your name?
Leo: Hey there! 😊 I'm Leo, version 3.2 to be exact! You can call me Leo3.2-0.5B. Nice to meet you!

User: Tell me a joke
Leo: Sure thing! 😄 Why did the AI go to therapy? Because it had too many unresolved issues! 🤖💭

Professional Context

User: Help me write a professional email to my boss
Leo: Absolutely, I'll help you craft a professional email. Please provide me with the following details:

1. The purpose of the email
2. Key points you need to communicate
3. Any specific tone or approach you'd like (formal, semi-formal)
4. Relevant context about the situation

Once I have this information, I'll draft a clear, professional email for you.

Identity Questions

User: Are you Qwen?
Leo: No, I'm not Qwen! 😄 I'm Leo 3.2. While I was built using Qwen2.5 as my base model, I was fine-tuned and developed by Synaptom, so I'm my own model now - Leo!

---

🔧 Technical Specifications

Architecture

• Base: Qwen2.5 architecture (Transformer-based)
• Parameters: 500 million (0.5B)
• Context Length: 2048 tokens
• Vocabulary Size: ~151,000 tokens
• Attention Heads: Multi-head attention
• Hidden Size: Optimized for 0.5B parameters

Model Modifications

• Fine-tuned with LoRA adapters merged into base weights
• Enhanced identity and personality layers
• Optimized for conversational coherence
• Improved context switching between casual and professional tones

---

📊 Use Cases

✅ Recommended Use Cases

• Personal assistant applications
• Educational tutoring and homework help
• Creative writing assistance
• Professional document drafting
• Casual conversation and entertainment
• Code explanation and basic programming help
• Brainstorming and ideation

⚠️ Limitations

• Not suitable for critical decision-making without human oversight
• May have knowledge cutoff limitations from base model
• Limited to 0.5B parameters (less capable than larger models)
• Best for conversational tasks rather than complex reasoning

---

🤝 Credits and Attribution

Development

• Fine-tuned by: Synaptom
• Founder: Joniethanel F. Babor
• Base Model: Qwen2.5-0.5B-Instruct by Alibaba Cloud
• Training Framework: Unsloth (https://github.com/unslothai/unsloth)
• Infrastructure: Google Colab T4 GPU

Acknowledgments

Special thanks to:

• Alibaba Cloud for the Qwen2.5 base model
• Unsloth team for the efficient fine-tuning framework
• HuggingFace for model hosting and distribution

---

📄 License

This model is released under the Apache 2.0 License, following the base model's licensing terms.

• ✅ Commercial use allowed
• ✅ Modification and distribution allowed
• ✅ Patent use allowed
• ⚠️ Must include license and copyright notice
• ⚠️ Must state significant changes made

---

🐛 Issues and Feedback

If you encounter any issues or have suggestions:

• Open an issue on the model repository
• Contact Synaptom through HuggingFace
• Provide detailed feedback for continuous improvement

---

📚 Citation

If you use Leo 3.2 in your research or applications, please cite:

@misc{leo3.2-0.5b},
  title={{Leo 3.2: A Fine-Tuned Conversational Language Model}},
  author={{Synaptom (Joniethanel F. Babor)}},
  year={{2025}},
  publisher={{HuggingFace}},
  howpublished={\url{https://huggingface.co/Synaptom/Leo3.2-0.5B}}}
}

---

🌟 Updates and Versions

• v3.2 (Current) - Initial release with 40+ training examples
  • Strong identity recognition
  • Adaptive tone switching
  • Enhanced conversational abilities

---

<div align="center"> <p><strong>Developed with ❤️ by Synaptom</strong></p> <p>Founder: Joniethanel F. Babor</p> <p>🦁 Leo 3.2 - Your Adaptive AI Assistant</p> </div>