minkarr 0.1.6

A minimal implementation of KaRR knowledge assessment method for Large Language Models (LLMs)

Statistical Knowledge Assessment for Large Language Models

A minimal implementation of KaRR knowledge assessment method from the following paper:

Statistical Knowledge Assessment for Large Language Models,
Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Zhifang Sui, Lei Li
arXiv preprint (arxiv_version)

This is a fork of the official implementation released by the authors.

How to use?

First, create a new virtual environment, then install Pytorch-CUDA, and finally install minkarr using the command:

pip install minkarr

Here is a simple example of how to quantify the knowledge of a fact by an LLM using KaRR

from minkarr import KaRR
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "gpt2"
device = "cuda"
model = AutoModelForCausalLM.from_pretrained(model_name).cuda()
tokenizer = AutoTokenizer.from_pretrained(model_name)

karr = KaRR(model, tokenizer, device)

# Testing the fact: (France, capital, Paris)
# You can find other facts by looking into Wikidata
fact = ("Q142", "P36", "Q90")

karr, does_know = karr.compute(fact)
print("Fact %s" % str(fact))
print("KaRR = %s" % karr)
ans = "Yes" if does_know else "No"
print("According to KaRR, does the model knows this fact? Answer: %s" % ans)
# Output:
# KaRR = 3.338972442145268
# According to KaRR, does the model knows this fact? Answer: No

Difference with original repo

• Easy-to-use
• Cleaner code
• Minimalistic implementation: I kept only the portion of the code needed to compute KaRR and removed the rest
• This implementation can compute KaRR on a single fact (the original implementation went through all facts)

Change storage location

Set the environment variable STORAGE_FOLDER to choose where to store the data that MinKaRR downloads.

Citation

Cite the original authors using:

@misc{dong2023statistical,
      title={Statistical Knowledge Assessment for Large Language Models}, 
      author={Qingxiu Dong and Jingjing Xu and Lingpeng Kong and Zhifang Sui and Lei Li},
      year={2023},
      journal = {Proceedings of NeurIPS},
}