If I want to fine tune on Chinese data, how much data volume and GPU resources do I need?
Sunting78 opened this issue · comments
thanks
I think you can do some fine tuning based on ChartAst-S. The simplest way is:
- Generate Chinese chart-table pairs by yourself and perform stage-1 training
- Refer to the QA generation process in the article to generate QA data for the data in step 1 and perform training.
It is difficult to say the specific requirements, because the process in step 1 can actually generate a lot of samples, so I suggest you start with 50,000 pairs. This process takes about half a day on an 8-card A100
Thanks a lot for your reply.