8 real, unedited runs on fresh GSM8K test questions. Expand any question to see all candidates, the calculator check, and the verifier's pick.
#### <number> line the training data uses, with a careful fallback
for responses that state the answer in plain prose instead. These examples come from an earlier evaluation run, so the wording may differ from the final adapter, which writes <<...>> steps and ends with #### N. See the
overview page for the training story and the final results.