Improving Diversity of Commonsense Generation by Large Language Models via In-Context Learning

---

Introduction

This document provides the code and explanations of the 2024 EMNLP Findings for the paper Improving Diversity of Commonsense Generation by Large Language Models via In-Context Learning

Requirements

Before running the code, ensure you have all the necessary dependencies installed.

conda create -n icd python=3.10
conda activate icd
pip install -r requirements.txt

Please

1. put the dataset folder from the downloaded data to the /.
2. data folder from the downloaded data to extrinic folder.

Also, you need to have the tokens from OpenAI and Huggingface to run the code.

Training & Evaluation

We have already uploaded the results and the generated sentences from our experiments into the folder so you could run with our provided evaluation code eval_accuracy.py in eval folder.

To run the ICD method on GPT3.5-Turbo, you need to run Intrinic/candidate.py to generate the default and diversified sentences and then run Intrinic/icd.py.

To run our experiments on the BART, set the default --method_name moree in main.py then run.

To run our ICD experiments on Vicuna, we follow the advice from vicuna authors and launch a RESTful API Server.

1. cd to Intrinic/Vicuna folder
2. launch the controller python3 -m fastchat.serve.controller
3. launch the model worker python3 -m fastchat.serve.model_worker --model-path lmsys/vicuna-13b-v1.5
4. launch the RESTful API server python3 -m fastchat.serve.openai_api_server --host localhost --port 8000

To try finetune on Vicuna-13b-v1.5, please run the data preprocessing script Intrinic/Vicuna/data/finetune_format.py and then run the Intrinic/Vicuna/train.sh. The generation code for finetune is in Intrinic/Vicuna/generate.py

python main.py