This repo contains the source code for reproducing the experimental results reported in paper: "Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space", which is accepted to Neurips 2024 (Arxiv link: https://arxiv.org/abs/2405.13845).
Below are the step-by-step guideline for setting up the experiment environment:
(a) Use the environment_llama2_mistral_mixtral.yml
file to create an anaconda environment for all the experiments with Llama-2-13B
, Llama-2-70B
, Mistral-7B
, Mixtral-8x7B
and Mixtral-8x22B
.
(b) Replace anaconda3/envs/{your_env_name_for_llama2_mistral_mixtral}/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.py
with modeling_llama2.py
in folder huggingface_replacement
(change the file name to modeling_llama.py
).
(c) Replace anaconda3/envs/{your_env_name_for_llama2_mistral_mixtral}/lib/python3.10/site-packages/transformers/models/mistral/modeling_mistral.py
with modeling_mistral.py
in folder huggingface_replacement
.
(d) Replace anaconda3/envs/{your_env_name_for_llama2_mistral_mixtral}/lib/python3.10/site-packages/transformers/models/mixtral/modeling_mixtral.py
with modeling_mixtral.py
in folder huggingface_replacement
.
(e) Use the environment_llama3.yml
file to create an anaconda environment for all the experiments with Llama-3-8B
and Llama-3-70B
.
(f) Replace anaconda3/envs/{your_env_name_for_llama3}/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.py
with modeling_llama3.py attached
(change the file name to modeling_llama.py
).
Please read the detailed step-by-step guideline inside folder experiment_code
to generate the experimental results.
If you find semantic density useful, please cite it using the following BibTeX entry:
@inproceedings{qiu2024semantic,
title={Semantic Density: Uncertainty Quantification for Large Language Models through Confidence Measurement in Semantic Space},
author={Qiu, Xin and Miikkulainen, Risto},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
}