ChatGPT Quantitative

Experiment with Summarizing on quantitative data (real numbers).

report = “”"

County A: Min 6.8, P10 is 12.1, P50 is 23.6, P90 is 36.79, Max 42.1.\

County B: Min 8.82, P10 is 11.7, P50 is 27.6, P90 is 42.53, Max 48.8. \

County C: Min 12.56, P10 is 16.8, P50 is 38.9, P90 45.21, Max 60.74. \

The area generates more heat than expected.


Prompt as follows:

prompt = f"“”

Range is defined as the difference between minimum and maximum.\

Your task is to generate a short summary of \

report focusing on the range and precentiles.

Summarize the report below, delimited by triple


Review: {report}


response = get_completion(prompt)



The report presents data on three counties, A, B, and C, focusing on their minimum, maximum, and percentiles (P10, P50, and P90). County A has a range of 35.3 (from 6.8 to 42.1), County B has a range of 40.98 (from 8.82 to 48.8), and County C has a range of 48.18 (from 12.56 to 60.74). The percentiles show that County C has the highest temperatures, with a P90 of 45.21, while County A has the lowest temperatures, with a P10 of 12.1.

The report concludes that the area generates more heat than expected.

Observation - it looks good but the County B range value (as defined) is off by 1.0.

Q: is this a limitation of gpt-3.5-turbo or LLM when working with real numbers?

Much appreciated for clarification and comments. Pat

This is a good experiment, and another case where I can see that the models are not necessarily great at math with common prompts.

To make it work, you’ll have to guide the model with a very detailed, step-by-step process. Or may be even try to use this at the begining: “Resolve the exercise below step by step”. Try that and lets see if it works better.

1 Like

Much appreciated the suggestion, Juan.

Observations -

  • Summarizing is really good
  • Computing correctly is a challenge (GPT not a calculator)

Perhaps the safe practice is to do the required calculation elsewhere (just as percentiles) and use ChatGPT to summarize.