blue-dragon-3b-v1

license: llama3 base_model: unsloth/Llama-3.2-3B-Instruct-bnb-4bit tags:

• lora
• sft
• unsloth
• transformers
• text-generation
• reasoning
• project-blue-dragon

---

🐉 Blue Dragon v1: An Exploration of Emergent Adaptive Reasoning

Blue Dragon v1 is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct, developed as part of Project Blue Dragon.

This project began as a solo experiment on a single NVIDIA RTX 3060 to teach a 3B model a step-by-step reasoning process. Along the way, an unexpected and powerful capability emerged: the model learned to treat reasoning as a cognitive tool, and discovered on its own when to use it.

This is a proof-of-concept for a new path in AI development: creating models that are not just more powerful, but more intelligent and efficient in how they apply their power.

---

💡 The Core Ability: Knowing When to "Think"

The model operates in two distinct modes, choosing the correct one based on the user's query.

1. Conversational Mode (Tool Not Needed)

When presented with a simple, conversational prompt, the model responds directly and naturally.

User:

Hi

Blue Dragon v1:

Hello! It's nice to meet you. Is there something I can help you with?

2. Reasoning Mode (Tool Activated)

When presented with a problem that requires logic or calculation, the model activates its fine-tuned "thinking tool" to break down the problem and show its work.

User:

If a pizza is cut into 8 slices and you eat 3, what fraction of the pizza remains?

Blue Dragon v1:

Let me solve this step by step:

<thinking> First, I need to understand what the question is asking. The question is: If a pizza is cut into 8 slices and you eat 3, what fraction of the pizza remains?

I'll break this down:

1. Identify the key information and what we're solving for
2. Set up the equation or approach
3. Solve step by step
4. Verify my answer

Let me work through this systematically: </thinking>

<answer> 5/8 </answer>

---

🔬 The Discovery: A Happy Accident

This adaptive behavior was an emergent property, not an explicitly programmed goal. My working hypothesis is that this is a result of two key factors:

• Focused Data: The fine-tuning dataset (~2000 examples) was highly specialized for reasoning. This taught the model to associate specific problem-solving patterns with the use of the "thinking tool."
• LoRA Architecture: By using LoRA, we updated only a tiny fraction of the model's weights. This created a specialized "reasoning pathway" without overwriting the model's vast, pre-existing conversational abilities trained by Meta.

---

🚀 How to Use

This is a LoRA adapter. To use it, first load the base model with Unsloth, then apply the adapter.

from unsloth import FastLanguageModel
import torch

# 1. Load the base Llama 3.2 3B model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Llama-3.2-3B-Instruct-bnb-4bit",
    load_in_4bit = True,
)

# 2. Load the Blue Dragon LoRA adapter from Hugging Face
model.load_adapter("noobezlol/blue-dragon-3b-v1")

# 3. Run Inference
# (See the inference.py script in the GitHub repo for a full chat example)

➤ For the full story, complete source code, and methodology, please visit the official Project Blue Dragon GitHub Repository.

🙏 Attribution

    Base Model : This project stands the shoulders of Llama 3.2-3b a powerful model by Meta. This work adheres to the Llama 3.2 Community License Agreement.
    Fine-Tuning Library: Made possible by the incredible efficiency of the Unsloth library.