2/ We started with simple prompting — tell the model what you're looking for and how to format its responses. This catches a lot of the low hanging fruit (~49 F1), but still needs refinement.

→ View original post on X — @snorkelai