Nova 0.5 E3 7B

🚀 Nova 0.5 e3

This model is capable of solving simple and complex math problems, demonstrating an interesting emergent reasoning property.

✨ Features

Reasoning Ability

This model shows an interesting emergent property. When asked a simple prompt like What is 4 x 2?, it'll simply answer 8. After some tests, it was found that when presented with a more complicated problem like Solve x^2 + x = 8., it begins to reason:

<|begin_of_thought|>
**Analysis:**
The equation is a quadratic equation in standard form \(x^2 + x - 8 = 0\). Factoring directly seems challenging, so substitution and factoring of a perfect square trinomial is attempted.

**First:**
Rewrite the equation by adding 1 to both sides:
\(x^2 + x + 1 = 9\).

**Second:**
Recognize that \(x^2 + x + 1\) can be written as \((x + \frac{1}{2})^2 - (\frac{1}{2})^2 + 1\), simplifying to \((x + \frac{1}{2})^2 - \frac{1}{4} + 1 = (x + \frac{1}{2})^2 + \frac{3}{4}\).

**Next:**
Set this equal to 9:
\((x + \frac{1}{2})^2 + \frac{3}{4} = 9\).

**Reflection:**
Subtract \(\frac{3}{4}\) from both sides:
\((x + \frac{1}{2})^2 = 9 - \frac{3}{4} = \frac{36}{4} - \frac{3}{4} = \frac{33}{4}\).

**Finally:**
Take the square root of both sides:
\(x + \frac{1}{2} = \pm \sqrt{\frac{33}{4}} = \pm \frac{\sqrt{33}}{2}\).

**Summarizing:**
Solving for \(x\):
\(x = -\frac{1}{2} \pm \frac{\sqrt{33}}{2} = \frac{-1 \pm \sqrt{33}}{2}\).
The solutions are:
\(\boxed{\frac{-1 + \sqrt{33}}{2}, \frac{-1 - \sqrt{33}}{2}}\).
<|end_of_thought|>

<|begin_of_solution|>
**Solution:**
Starting with \(x^2 + x = 8\), rewrite it as \(x^2 + x + 1 = 9\). Completing the square on the left side:

1. Rewrite \(x^2 + x + 1\) as \((x + \frac{1}{2})^2 - (\frac{1}{2})^2 + 1 = (x + \frac{1}{2})^2 - \frac{1}{4} + 1 = (x + \frac{1}{2})^2 + \frac{3}{4}\).
2. Set equal to 9: \((x + \frac{1}{2})^2 + \frac{3}{4} = 9\).
3. Subtract \(\frac{3}{4}\): \((x + \frac{1}{2})^2 = \frac{33}{4}\).
4. Take square roots: \(x + \frac{1}{2} = \pm \frac{\sqrt{33}}{2}\).
5. Solve for \(x\): \(x = -\frac{1}{2} \pm \frac{\sqrt{33}}{2}\).

Thus, the solutions are \(\boxed{\frac{-1 + \sqrt{33}}{2}, \frac{-1 - \sqrt{33}}{2}}\).
<|end_of_solution|>

As shown, the model can reason and get the correct answer. Although this reasoning ability is not as good as that of Nova - 0.5 - r1 - 7B, it still shows promise.

Accident Discovery

The ability to reason was achieved by complete accident. A reasoning dataset was accidentally mixed in with other simpler datasets, which taught the model to reason only during more complex, multi - stepped prompts.

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic Usage

Below is a basic example to load and run Nova 0.5 e3 using Python and the Hugging Face transformers library. Make sure you have the required dependencies installed (transformers, torch, etc.).

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the tokenizer and model from Hugging Face
model_name = "oscar128372/Nova-0.5-e3-7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Move to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Setup ChatML prompt
chatml_prompt = """
<|im_start|>system
{}<|im_end|>
<|im_start|>user
{}<|im_end|>
<|im_start|>assistant
"""

# Example system prompt
system_prompt = "You are a helpful assistant."

# Example prompt
prompt = "Solve x^2 + x = 8."

# Tokenize input
inputs = tokenizer(
[
    chatml_prompt.format(
      system_prompt,
      prompt
    )
], return_tensors="pt").to(device)

# Generate response
outputs = model.generate(
    **inputs,
    max_length=1024,  # Keep this high for reasoning, else, keep it low.
)

# Decode and print the result
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

# Expected output: A reasoned solution, e.g., "x = (-1 ± √33)/2"

📚 Documentation

Precautions before Use

• Loading in 4 - bit: You cannot load it in 4 bit. Loading it in 4 bit removes the reasoning ability entirely and turns it into an actual base model.
• ChatML Template: Use the ChatML template:

<|im_start|>system
{}<|im_end|>
<|im_start|>user
{}<|im_end|>
<|im_start|>assistant
{}

⚠️ Important Note

I have only tested the model using "You are a helpful assistant." system prompt, so other system prompts may produce incorrect or unexpected results.

📄 License

The model is licensed under the apache - 2.0 license.

What's next?

There will be no e4 in the future. The next step is 1.0, and maybe 1.0 - r1? Keep an eye out for any new reasoning models in the future. :)

Additional Information