That's exactly what happened to me when I tried to get the open-source models to extract a CSV from textual data with a lot of yes/no fields; i.e., the model fo

That's exactly what happened to me when I tried to get the open-source models to extract a CSV from textual data with a lot of yes/no fields; i.e., the model forgot the column order and started confusing the cell values. I found I had to use more powerful models like Mistral Large or ChatGPT. So I think that is a valid thing to worry about with smaller models, but maybe less of a concern with larger ones.