I have been researching Language Models that can work with tabular data. My main goal is to have a model to answer simple questions about my data. An example is having household sales data and asking simple questions like "What was the average sales during the last 2 months?". One of the best models I have found so far is TAPAS. However, it has limitations regarding the size of tabular data. My data size is approximately 1 million rows with 10 columns. Is there a robust model that can perform the mentioned task or is there an alternative approach to this problem?
Language models will experience difficulties processing in an analytical way tabular data and answering complex questions about it because of how they work (just predicting the next token). Maybe you can achieve some interesting results with really big models but an approach that seems more promising to me is to pass the basic structure of a dataset to the models, like describing what a row and what each column represents, and asknig the model to write a SQL query that answers your questions.
Example: "My dataset consists of household sales data. Each row is a sale and there are 3 columns: SaleID, Price, Year. Write a SQL query that generates the most expensive house sold by year."
Yes. Use GPT4 to generate code to print output, then it can read and interpret the output. This is already a product offered by a few companies that people use.
I have done exactly what you are describing on datasets that, coincidentally, also have almost exactly 1 million rows and exactly 10 columns.