Benchmarks achieved with benchmark hunting ?

[Greatz08]

Asking this here because gpt 3.5 as far as i remember is trained on much bigger dataset and can generate much large text . I do agree it is old now but still i was very satisfied with its bigger response and way of detailed explanation compared to today's gen smaller 8b models.
So i wonder if these benchmarks are legit because i have used countless small llm models which shows benchmark on par with gpt 4 but cant do sometimes even many basic things because it is quite practical that with such low dataset and only some billion parameters by no mean are it is enough to satisfy in same way as bigger models like llama,gpt,claud can and only possible way to achieve benchmarks in today's world is through training for benchmark hunting which is not healthy for open source development in my opinion but i have to admit this is the case for most unfortunately.