These are full episodes from the text-to-SQL runs in our AsyncGRPO benchmark post. The same LoRA recipe is trained in
TRL, prime-rl and SkyRL.
In each episode the model reads a SQLite schema and a question. It can run up to 5 read-only queries with execute_sql_query, then gives one SQL answer. The answer is graded by running it and comparing its result with the gold query's result.
30 tasks × 13 checkpoints
correctwrongno answernumber = SQL callsa query erroredClick a cell to open that episode. Arrow keys move across tasks and checkpoints.