ChatGPT with its own model

Everyone is talking about ChatGPT and I have to admit that I use it very intensively:

  • Writing product descriptions
  • Find ideas for names
  • Generate proposals for plans
  • Help with writing e-mails
  • Help with writing code
  • When writing blog posts
  • and much more

These are all relatively simple tasks that always run in the existing model. I am ready for another step.

Mapping of data – task definition

The task is relatively easy to understand: A list of product categories from an e-commerce store is to be mapped to the Google Shopping product categories. The source list includes about 80 categories, while Goolge’s includes thousands.

First attempt: ChatGPT

In my first attempt, I simply copied the list into ChatGPT and told him to copy the corresponding product categories from Google Shopping. He said he only had data up to 2021. At least it was positive, ChatGPT understood the task and did something.

But unfortunately not what I wanted. There was always a new ID behind it, but a simple cross-check quickly showed that this Google product category either does not exist or the assignment is simply incorrect.

What is not entirely clear, however, is whether the data is outdated or the mapping is so bad.

Second attempt: Train your own model

It should actually be possible to train the model with the current data. I finally find someone who has done something similar and loaded all the US tax data into the model. I load the model with the new data and try again.

Unfortunately, again without a satisfactory result. It may well be that I have not done everything correctly.

Mission Failed – but by hand

Since a few hours have passed in the meantime, I decide to do the mapping by hand. It’s not that much data now and investing more hours or days would be a bit disproportionate.

Conclusion – ChatGPT

Absolutely terrific and saves me hours of work, but the tools based on it are still in their infancy. There is still a lot to do here. I already have a lot of ideas, but unfortunately I simply don’t have the time to read up on them quickly enough.

And here is the article: Developing TaxGPT using OpenAI GPT and Chroma