To improve the MGSM benchmark, we corrected two erroneous English questions and rephrased others to remove ambiguity. We then used Gemini to retranslate all questions and subsequently used Gemini to verify that every question in the benchmark is now answerable. - View it on GitHub