TL;DR

• This is a simple math reasoning test: first filter (pick odd numbers), then add them, then judge parity (odd/even).
• Good for observing: whether the model follows steps, whether it makes intermediate calculation errors, and whether it stays consistent.
• In production: require output of intermediate results at each step (list, summation process, final conclusion) for easier evaluation.

Background

This prompt tests an LLM's mathematical capabilities by asking it to check whether the odd numbers in a group add up to an even number. It also nudges the model to break the problem into steps.

How to Apply

Treat this prompt as a "step-by-step execution" template:

1. Identify (filter odd numbers)
2. Compute (sum them)
3. Verify (check if result is odd/even)

To migrate this to business calculation tasks, the key is: make each step a verifiable sub-task, and require intermediate results in the output.

How to Iterate

1. Structured output: require a small table numbers / is_odd / included / running_sum
2. Add self-check: repeat the sum at the end, or re-compute in a different order
3. Control temperature (math tasks generally benefit from lower temperature for consistency)
4. Adversarial testing: make the list longer, add invalid characters or more complex formatting, test robustness

Self-check Rubric

• Did it correctly identify which numbers are odd?
• Is the sum correct? Any intermediate step errors?
• Is the final parity judgment consistent with the sum?
• Does the output include verifiable intermediate results?

Practice

Exercise: apply the same template to these two task types:

• From a mixed input, extract numbers meeting a condition (e.g., divisible by 3), then sum
• Given a list of dates, filter out weekends, then count

Prompt

The odd numbers in this group add up to an even number: 15, 32, 5, 13, 82, 7, 1.
Solve by breaking the problem into steps. First, identify the odd numbers, add them, and indicate whether the result is odd or even.

Code / API

OpenAI (Python)

from openai import OpenAI

client = OpenAI()

response = client.chat.completions.create(
    model="gpt-4",
    messages=[
        {
            "role": "user",
            "content": "The odd numbers in this group add up to an even number: 15, 32, 5, 13, 82, 7, 1.\nSolve by breaking the problem into steps. First, identify the odd numbers, add them, and indicate whether the result is odd or even.",
        }
    ],
    temperature=1,
    max_tokens=256,
    top_p=1,
    frequency_penalty=0,
    presence_penalty=0,
)

Fireworks (Python)

import fireworks.client

fireworks.client.api_key = "<FIREWORKS_API_KEY>"

completion = fireworks.client.ChatCompletion.create(
    model="accounts/fireworks/models/mixtral-8x7b-instruct",
    messages=[
        {
            "role": "user",
            "content": "The odd numbers in this group add up to an even number: 15, 32, 5, 13, 82, 7, 1.\nSolve by breaking the problem into steps. First, identify the odd numbers, add them, and indicate whether the result is odd or even.",
        }
    ],
    stop=["<|im_start|>", "<|im_end|>", "<|endoftext|>"],
    stream=True,
    n=1,
    top_p=1,
    top_k=40,
    presence_penalty=0,
    frequency_penalty=0,
    prompt_truncate_len=1024,
    context_length_exceeded_behavior="truncate",
    temperature=0.9,
    max_tokens=4000,
)

Reference

• Sparks of Artificial General Intelligence: Early experiments with GPT-4 (13 April 2023)