ChatGPT and CSV Files: Why LLMs Struggle with Structured Data

If you’ve tried using ChatGPT to create or read CSV files, you may have encountered some frustrations. While large language models (LLMs) like ChatGPT are incredibly powerful for generating text and handling natural language, they have limitations when it comes to working with structured data like CSVs. In my recent experience, despite their impressive abilities, Chat GPT is not, in its basic format, the tool for structured data like CSV.

I created a CSV using ChatGPT for a batch-generation project. The task was relatively simple: I needed the model to generate text chunks from a prompt and split them into different categories, with each line containing at most 20 words, and follow a specific formatting that I provided. I provided precise instructions on how to structure the information. Despite the simplicity of the task, ChatGPT repeatedly made errors—mixing up lines, miscounting entries, and even adding extra blank lines and categories. Correcting it led to more mistakes, and soon, the context of the exchange got muddled. The point was to make the process simpler and more straight-forward but Chat GPT consistently fails. You can get through it with corrections, but it is not ideal.

So, why is this happening?

LLMs like ChatGPT are trained to process text, not structured data. They excel at generating language and making sense of natural language queries. They struggle with understanding the inherent structure of CSV files. Commas in CSVs are used as delimiters, which can confuse an LLM, interpreting them as part of the text rather than the boundaries between data fields. (1,2)

LLMs aren’t naturally equipped to process rows and columns in the way that tools like spreadsheets or databases do. It’s not that the model can’t generate structured data—it can. But without specific prompts or instructions, it doesn’t inherently understand how to organize that data in a way that matches the structure of a CSV.

If you need to generate a CSV or work with structured data, try using a tool like Python to do the heavy lifting. You can ask ChatGPT to generate Python code that creates the CSV, ensuring the data is formatted correctly. This works especially well to generate lists or structured data, such as product names, prices, or batch-generation prompts. This isn’t difficult, when you are teaching it, ask it to generate and USE the code it makes. It’s literally a couple more minutes of creating the GPT and it will stop making errors. 

When using ChatGPT for CSV-related tasks, keep these tips in mind:

1. **Be specific with your prompts**: The more specific you are, the better your results will be. For example, clearly define the column names and data types if you’re generating a CSV.

2. **Consider preprocessing your data**: If you need to feed CSV data into ChatGPT, converting it into a more natural language format is often more effective. This gives the model the context it needs to analyze and generate insights.

While LLMs like ChatGPT can be incredibly useful for many tasks, they’re not the best option for reading and generating structured data like CSV files. Using them alongside more specialized tools is your best bet for effective, efficient results.

I created two GPTs this week to test this:

1. A Batch Generation tool for Ideogram that I created with Chat GPT. It tends to make errors and need corrections.
2. A Batch Generation tool for Ideogram that I created with Chat GPT and Python. It has not made any errors in my limited testing.

If you are testing batch generation in Ideogram like we discussed last week, give them a try. (The funny thing is ChatGPT is like a human when it names its CSV variations…. Names.csv –> FinalNames.csv –> FinalNames2.csv –> FinalFinalNames2.csv)