oscar128372/Nova-0.5-e3-7B

Nova 0.5 e3

This model shows an interesting emergent property.

If asked a simple prompts like What is 4 x 2?, it'll say 8. Simple, right?

Well, after some tests. I found out that when asked a problem that is more complicated like Solve x^2 + x = 8., something interesting happens.

<|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 you can see, it begins to reason. Not only that, it got it correct!

Of course, this reasoning is not as good as Nova-0.5-r1-7B, but it still shows promise!

How was this achieved?

Believe it or not, this was by complete accident. I accidentally mixed in a reasoning dataset along with other simpler datasets! Oops...

Turns out, the reasoning dataset taught the model to only reason during more complex, multi-stepped prompts.

What you NEED to know before using:

Number 1, you cannot load it in 4 bit. I was so confused on why it wasn't working. Loading it in 4 bit just removes the reasoning entirely and turns it into an actual base model, so if you don't want reasoning, that works, I guess? I don't really know why.

Number 2, use the ChatML template:

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

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

How to run the model?

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"

What's next?

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