Harder maths problems, checked by code

For my MSc thesis (UvA, 2025), a teacher model rewrote maths problems into harder variants with a Python solution.

A verifier ran that code and kept only variants that ran, gave the right answer and were new to the model: 6,213 problems from 7,460 originals, harder than the hand-made GSM-Symbolic benchmark. In a manual sample, 81 of 100 were fully correct. Honest result: fine-tuning on the new data gave no significant improvement.