H3 - synthetic data

useful when there are insufficient example data

useful for:

• increasing training volume
• increasing coverage
• rare cases
• cost

synthetic generation might help create additional diverse examples of the rare scenarios, called augmentation

garbage in, garbage out applies

diversity of synthetic data matters

synthetic generation can reproduce or amplify patterns and biases present in:

• the generator
• source examples
• prompt
• selection process

evaluation should compare whether synthetic augmentation actually improves performance on representative real evaluation data