Understanding GPT-3 and multi-language

I’ve some basic understanding problems of gpt-3 and multi-language.

My understanding is, that gpt-3 ist manly trainied on English texts like English literature and English websites.

If I make a query with some German text, the results of gpt-3 are also very good and make sense in the most use cases.

Why is gpt-3 also able to reply to German prompts?

Is the architecture of gpt-3 so universal, that language doesn’t matter, or is the amount of German texts gpt-3 is trained with also high enough to make this very good completitions?

In other words: Can I train a model with only English text, and it will also work in other languages, or do I need the model to train in the target promp language?


1 Like

Really interesting questions. Regarding the quote I feel it could be both.

At least with the web version https://chat.openai.com/chat, it is possible. We started using it to translate the dialogs in the game we are working on to German, Spanish, French, Italian, Dutch, Portuguese from Portugal, and Portuguese from Brazil, and it works really well. Much better than DeepL and Google Translate, which we used before through the API.
I am in the process of creating a small Python application to do our translation using the OpenAI API and hope to achieve similar results to the web version.