Title

From Values to Opinions: Predicting Human Behaviors and Stances Using Value-Injected Large Language Models

Overview

This code provides the implementation of "Value Injection Method (VIM)".

How to run

In src folder, we make bash script file to train and inference, and evaluate for Value Injection Method (VIM).

Please move the data folder into the src before running.

In this code, train, reference, and performance calculation are conducted for 128 groups (28 country groups, 100 social groups) obtained from the European Social Survey.

• train.sh : a bash script file to train LLaMA-7B with VIM

bash train.sh

• inference.sh: a bash script file to inference LLaMA-7B with VIM

bash inference.sh

• calculation.sh : a bash script file to calculate the evaluation task 1, 2, 3, 4

bash calculation.sh

• preprocessing.py : a python file to preprocessing Touche23-ValueEval dataset

python preprocessing.py

---

Train

This code train VIM, VIM_AG, VIM_QA method.

• VIM
• VIM_AG
• VIM_QA

---

Model

• LLaMA-7B (7 billion parameters)

---

Evaluation Task

1. Portrait Values Questionnaire (PVQ)
2. Touché23-ValueEval dataset (Argument Generation)
3. VALUENET
4. European Social Survey (ESS)

---

Metric

• Normalized Mean Squared Error (NMSE)

---

Dataset

We use these public dataset ...

• Touché23-ValueEval dataset : https://touche.webis.de/semeval23/touche23-web/index.html
• VALUENET dataset : https://touche.webis.de/semeval23/touche23-web/index.html
• ESS : https://www.europeansocialsurvey.org/

---

Tested Environment

• python 3.9.7
• numpy 1.22.4
• pandas 1.3.3
• scikit-learn 1.0
• torch 1.13.1
• tokenizers 0.13.2
• transformers 4.27.0.dev0
• peft 0.3.0.dev0

---