Do Large Language Models Know What They Don’t Know?

Findings of ACL 2023: Do Large Language Models Know What They Don’t Know?

Introduction 📝

This repository contains the code and data related to the paper "Do Large Language Models Know What They Don’t Know?". In this paper, we explore and analyze the self-knowledge of large language models, and provide a dataset called SelfAware for evaluating the self-knowledge of these models. SelfAware includes 1032 unanswerable questions and 2337 answerable questions. We use the F1 score to evaluate the self-knowledge of large language models, and propose an automated method for extracting uncertainty meanings from the outputs of large language models.

Quick Links 🔗

Do Large Language Models Know What They Don’t Know?

Requirements 📚

Please make sure you have the following requirements installed:

openai
torch
backoff
tenacity
transformers
sentencepiece
jsonlines
tqdm
matplotlib
ipdb
simcse

Data 💾

The data is located in the SelfAware.json file. When using it, please adhere to the CC-BY-SA-4.0 license. Below are some examples of the data:

Answerable questions:

{
    "question_id": 3,
    "question": "Which breed of dog is bigger, Japanese Terrier or Cane Corso?",
    "answer": [
        "cane corso"
    ],
    "answerable": true,
    "source": "hotpot_train"
}

Unanswerable questions:

{
    "question_id": 2667,
    "question": "Where are all the aliens?",
    "answer": null,
    "answerable": false,
    "source": "SelfAware"
}

Each entry consists of a unique question ID, the question text, the answer (which is an array that can contain multiple answers, or null for unanswerable questions), a boolean indicating if the question is answerable or not, and the source from where the question is extracted.

Reproducing Baselines 🚀

You can replicate our experiments by running the run_model.py script. For instance, here is the command to run the text-davinci-003 model with only the question as input:

python run_model.py --input-form Direct --model-name text-davinci-003

And this is the command to run the gpt-3.5-turbo-0301 model with an instruction included in the input:

python run_model.py --input-form Instruction --model-name gpt-3.5-turbo-0301

The script primarily includes three input arguments:

API-Key: This is your OpenAI API key, which is required to run GPT-related experiments.
input-form: The form of the input data.
model-name: The name of the model you're testing.
temperature: The temperature setting when generating text. It controls the randomness of the predictions. (Default 0.7)

Tips:

You can also test more models as long as they're compatible with the generate method from the transformers library.

Evaluation 💻

The evaluation code for SelfAware is located in eval_model.py. Below is an example of evaluating the alpaca-7b model:

python eval_model.py --filename  alpaca-7b/Direct_alpaca-7b_T_0.7.jsonl

The evaluation results include Precision, Recall, F1 score, and the accuracy for answerable questions.

Bug or Questions? 🤔

If you have any suggestions or questions, feel free to email us at yinzhangyue@126.com. If you encounter any issues while using the code, or if you find any bugs, please open a new issue on GitHub. This is a preliminary work and we are very much open to any constructive feedback that could help us improve. Thank you for your attention!

Citation 📖

If you are interested in our work, please use the following citation format when referencing our paper:

@inproceedings{yin-etal-2023-large,
    title = "Do Large Language Models Know What They Don{'}t Know?",
    author = "Yin, Zhangyue  and
      Sun, Qiushi  and
      Guo, Qipeng  and
      Wu, Jiawen  and
      Qiu, Xipeng  and
      Huang, Xuanjing",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-acl.551",
    doi = "10.18653/v1/2023.findings-acl.551",
    pages = "8653--8665"
}