Getting from “this worked in a paper” to “this works in my training run” is often harder than it looks.
Papers are written to explain the idea and the result, not every detail needed to reproduce them. Turning that work into a reliable training recipe means filling in the gaps, testing assumptions, and understanding which choices actually matter.
River Recipes is an open-source initiative from River AI that turns leading research into practical, reproducible training recipes. Each recipe is tested, refined, and optimized by our team. You can run it as-is, or add your own data and adapt it to your use case.
Our first recipe follows our work on text-to-SQL: Achieving Frontier Text-to-SQL with the River API →