Open Emasoft opened 1 year ago
Hi Emasoft, thanks for your kind interest in our report.
Actually, our original intention is to maximize the translation performance of ChatGPT, and in this way, we give several pieces of advice in the report. Given the discussion on Reddit and the OpenAI official blog, we guess the multilingual processing ability improvement of GPT4 is mainly from the instruction tuning on the translated MMLU dataset.
Also, we sampled some test sets in our report to compare whether the proposed usage recipes could facilitate GPT4 either, and our conclusion is YES. More details will be updated in the report soon.
Welcome to try our proposed strategies in GPT4, and any feedbacks are welcome to post here or drop us an email.
Me (liangding.liam@gmail.com) and the first author (keqin.peng@buaa.edu.cn).
Interesting comparison between translation engines (GPT-4 beats all other translators): https://www.reddit.com/r/ChatGPT/comments/11t389v/wow_gpt4_beats_all_the_other_translators/