What's the SimpleQA score?
The tests you show are highly redundant, covering only a handful of domains. For example, math (MATH, CMATH & GSM8K), coding (EvalPlus, MultiPL-3 & MBPP), and STEM/academia (MMLU, MMLU-Pro, MMLU-Redux, GPQA, and SuperGPQA).
This is a big red flag. All models that have done this in the past have had very little broad knowledge and abilities for their size. And no publicly available tests do a better job of highlighting domain overfitting (usually math, coding, and STEM) more than the English and Chinese SimpleQA tests because they include full recall questions (non-multiple choice) across a broad spectrum of domains.
Plus Chinese models tend to retain broad Chinese knowledge and abilities, hence have high Chinese SimpleQA scores for their sized, because they're trying to make models that the general Chinese public can actually use. They only selectively overfit English test boosting data, resulting in high English MMLU scores, but rock bottom English SimpleQA scores.
I'm tired of testing these models since it's just one disappointment after another, so can you do me a favor and just publish the English & Chinese SimpleQA scores which I'm sure you ran so people can tell at a glance whether or not you overfit the math, coding, and STEM tests, and by how much?