3/ We then began data-centric prompt engineering, iteratively identifying model error modes and addressing them in the prompt where possible, either with added detail in the prompt or adding in-context examples (~68 F1).

→ View original post on X — @snorkelai