Number of question-answer pairs needed to build a private model for an enterprise

[sunilswain]

Hi,

I've been really enjoying how the community helps answer my questions, and I'm learning more about H2o.ai every day. Does anyone know how many question-answer pairs are needed to fine-tune a model effectively? I'm using h2ogpt-llama2-7b as the backbone, and I'm training the model on a single document that's about 20 pages long. Additionally, how many question-answer pairs would be required if I were to train the model on hundreds of documents of the same length?

Also please specify how many epochs are generally considered to be a good number?

Thanks