Comment by novaRom

Zero trust in benchmarks without opening model's training data. It's trivial to push results up with spoiled training data.