How much fine-tunning is necessary?

For those with experience, how much data do you have to feed OpenAI in order to get anything better than what davinci-002 can already produce?

Just as an experiment, I tried to feed 100+ examples in foreign language worth more than 15$ in tokens and the model started producing absolute gibberish compared to davinci-002. It feels like I would have to feed hundreds of dollars worht of data just so the model approaches the quality already produced by davinci-002. How is it with Enlgihs language?

I mean, is fine-tunning even cost-efficient considering that various fine-tuned tools are coming to the market bringing the costs per article down?

Seems people either don't know it or don't want to share?