For resolution criteria, it would be an Arxiv or conference paper on any model or any size (even small parameter counts) trained with gradient descent on Apple's M-series hardware. Adding a comparison paper for training a benchmark model on say an A100 vs an M-series would not count. It would only count if the primary model and focus of the paper was solely trained on M-series hardware, implying the topic of the paper was about modeling and not hardware comparison.