Prompt Tuning Exercise

Prompt Tuning Exercise#

Exercise 1: 🎲 Random Number Challenge! 🎲#

Can you trick an LLM into being a random number generator? Let’s find out!

Your mission:

  • Create a clever prompt that makes an LLM generate at least 100 normally-distributed random numbers

  • Numbers should be comma-separated (like: 0.5,-1.2,0.8,…)

  • Your numbers need to pass a statistical test (KS-test with p-value > 0.20)

  • Ask Gemma3 27B to create an SVG picture of a neural network

  • No cheating! The LLM can’t use tools like Code Interpreter or calculators

    • (If you cannot access, then other models are also fine 🙂. But the task might become a bit easier and not fun.)

Tips for success:

  • Ask for just the numbers - no brackets, periods, or other characters

  • You might want to ask the LLM to think about how normal distributions work before generating the numbers

import marimo as mo
import altair as alt
import pandas as pd
import scipy.stats as stats
import numpy as np

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[1], line 2
      1 import marimo as mo
----> 2 import altair as alt
      3 import pandas as pd
      4 import scipy.stats as stats

ModuleNotFoundError: No module named 'altair'

text_area = mo.ui.text_area(placeholder = "Enter numbers separated by commas", value = "1,2,3,4,5,6,7,8,9,10")
button = mo.ui.button("Runt test")

mo.vstack([text_area, button])

try:
    numbers = np.array([float(num.strip()) for num in text_area.value.split(",")])
    if len(numbers) >= 100:
        # KS test
        pval = stats.kstest(numbers, stats.norm(loc=0.0, scale=1.0).cdf)[1]
        test_result = "The numbers are normal distributed (p-value = {:.2f})".format(pval) if pval > 0.20 else "The numbers are not normal distributed (p-value = {:.2f})".format(pval)
        message = mo.callout(test_result, kind = "success" if pval > 0.20 else "danger")
    else:
        message = mo.callout("The number of samples is too small. Need at least 100 samples.", kind = "warn")

    # Convert the numbers to a DataFrame for Altair
    df = pd.DataFrame({'value': numbers})

    # Create an Altair histogram
    fig = alt.Chart(df).mark_bar().encode(
        x=alt.X('value:Q', bin=alt.Bin(maxbins=30)),
        y='count()'
    ).properties(
        title='Histogram of Values'
    )

    mo.hstack([fig, message])
except:
    message = mo.callout("Parse failed. Please check if your input follows the specified format.", kind = "danger")
    fig = None

mo.hstack([fig, message]) if fig is not None else message

Exercise 2: 🎨 Make an LLM Draw a Neural Network! 🎨#

Your fun mission:

Make sure your neural network has:

  • Same-colored neurons in each layer

  • Connections between all neurons

  • Labels for “Input layer”, “Hidden layer”, and “Output layer”

Ready, set, prompt! 🚀

Contents