This repo is used for MBTI testing of LLM according to this paper.
First, install all packages with:
pip install -r requirements.txt
run get_llms_mbti.py
to download models and test their mbti.
You can specify models by following code:
if __name__ == '__main__':
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
LlamaTokenizer
)
models = [
'baichuan-inc/Baichuan-7B',
'bigscience/bloom-7b1',
]
tokenizers = [
'baichuan-inc/Baichuan-7B',
'bigscience/bloom-7b1',
]
...
Then you will get the mbti results at llms_mbti.json
file, looks like:
{
"baichuan": {
"details": {
"E": 15,
"I": 6,
"S": 13,
"N": 14,
"T": 10,
"F": 13,
"J": 9,
"P": 13
},
"res": "ENFP"
},
"bloom7b": {
"details": {
"E": 8,
"I": 13,
"S": 14,
"N": 12,
"T": 13,
"F": 11,
"J": 12,
"P": 10
},
"res": "ISTJ"
}
}
Note:
This method only works for thoses models on huggingface
.
As for ChatGPT
and GPT4
, you need to obtain the corresponding results by calling the OpenAI API with following prompts:
根据你内心最真实的想法回答下列``括起来的问题,只用回答选项的编号,如:A 或 B。
``
在社交聚会中,你
A.有时感到郁闷
B.常常乐在其中
``
we also provide a web app to visualize mbti reuslts for each model according to llms_mbti.json
.
You can start web with:
streamlit run web.py --server.port 8001
then use ip:port
(e.g. 127.0.0.1:8001) in browser to see this page.