GPT-4o is a powerful LLM and free to use, however, when it comes to generating detailed and structured CSV files, GPT-4o has notable limitations.
Let's start with a straightforward task: generating 20 simple recipes. GPT-4o performs well here, creating a CSV with basic recipes that include just the name and a brief description.
Prompt
Give me 20 recipes in CSV format. Each recipe should include: name, ingredients, steps.
Result
So far, so good! GPT-4o handles simple requests with ease. But what happens when we crank up the complexity?
Now, let's up the ante. Generating 20 detailed recipes requires columns like name, ingredients, detailed steps, prep and cooking times, nutritional information, tags, user reviews, and a DALL-E prompt.
Prompt
Give me 20 recipes in CSV format. Each recipe should include the following details: name, ingredients, detailed steps, cookware required, prep time, cook time, 5 tags, nutrition information, a DALL-E prompt for the recipe, and 5 mocked user reviews. Ensure the recipe steps are detailed.
Result
I had to click "Continue generate" two times to complete this CSV. While the steps are somewhat more detailed, they still fall short, and the user reviews are brief. Let's push GPT-4o even further.
To push GPT-4o further, let's force a minimum content length in the prompt for some columns. Large language models (LLMs) often struggle with accurate length control, we provide additional hints only to indicate that we want more detailed responses.
Prompt