Gpt4 or even 3.5 is quite good at Finnish. Was there ever a benchmark against closed source models?