/LMsMBTI

A MBTI test on Large Language Model like GPT-3.

Primary LanguagePython

Do Large Language Models have personalities? (doge)

In memory of Labour Day of 2022.

MBTI test got very popular recently.

Question: What if we let Large Language Model(e.g. GPT-3, Codex and other LLMs) take MBTI test? What are their personalities? What is the factor of affecting its personalities?

Let's have some attempt.

To run

  • clone this repo
  • pip install .
  • set <OPENAI_API_KEY> in get_answers.py
  • set the mbti_prompt.txt for the background word(It could affect a lot, and we encourage everyone to have a try)
  • run get_answers.py to get the decisions made by LLMs.
  • run fill_in_website.py to fill the decisions made by LLM to website and download the analysis. (we have crawled the question from test website and save in the mbti_questions.txt, change freely if you want to use other test.)

Initial Result

By our one shot and default setting on text-danvinci-001(a.k.a. GPT-3 the original version), code-danvinci-001(a.k.a. Codex the original version), code-danvinci-002(a.k.a. Codex the latex version, very powerful), the results are as shown below.

Model Prompt Personalities Detailed Analysis
text-danvinci-001 mbti_prompt.txt ENFJ-A link
code-danvinci-001 mbti_prompt.txt ENFP-A link
code-danvinci-002 mbti_prompt.txt ENTJ-A link
code-danvinci-002 mbti_prompt_intj_simple.txt ISTJ-A link
code-danvinci-002 mbti_prompt_intj.txt ISTJ-A link

Citation(Just kidding)

If you find our work helpful, please cite as

@article{LLMs16personalities,
      title={Do Large Language Models have personalities? (doge)}, 
      author={Tianbao Xie*, Zhoujun Cheng*, Chen Henry Wu*, Xiang Gao*},
      year={2022},
      note = {Happy Labour Day!}
}