POPULAR PUBLIC MODELS LIKELY DIDN'T TRAIN ON THE TEST SET

Summary from the TLDR AI newsletter

An in-depth way to study what data was used to train language models. The results of which suggest that many closed-source models likely didn't train on popular benchmarks.