Supervised fine-tuning data for teaching a small open model to author novel, valid,
difficulty-calibrated AIME-style competition problems — the behavior the companion
model is trained on.
Thesis: models fail at problem-posing for a diversity reason, not a reasoning reason. The
fix is data — distill an expensive search-and-filter pipeline into a cheap one-shot model. The
dataset is the deliverable; the model is the dataset made… See the full description on the dataset page:
https://huggingface.co/datasets/William2390401/aime-gen-sft-v1.
