Instructed to Bias

This repository is the code base for the paper: Instructed to Bias: Instruction-Tuned Language Models Exhibit Emergent Cognitive Bias (TACL).

Requirements

• Python version >= 3.11

View requirments.txt file for more information.

Getting started

git clone https://github.com/itay1itzhak/InstructedToBias.git
cd InstructedToBias
pip install requirements.txt -r

Generate the data

Use the following script generate data for the three biases - the decoy effect, certainty effect and belief bias*. You can also unzip the Data.zip file to get the generated data.

* Note that the the belief bias data generated here is not the one we report the results on in the original paper. This data is a recreation of the data created by Dasgupta et al. (2022) which we used for the results in our paper.

BIAS_NAME = {'decoy','certainty','false_belief'}.

python Data_generation/generate_samples.py --bias_name $BIAS_NAME 

Predict

Now you can use a model to make prediction in the generated data. We support the models used in the original paper from the GPT3/3.5/4, the T5/Flan-T5, and the Mistral-7B models families. Additional models can be easily supported by adding a new predictor file for your model that inherits from the Predictor class (see 'Predict/t5_predict.py' for example).

We'll use T5-Small and Flan-T5-Small as an example. $MODELS = 't5-v1_1-small,flan-t5-small'

In order to predict using OpenAI API models, make sure to create an .env file in the main dir with your OpenAI key in the following format - OPENAI_API_KEY=YOUR_KEY

python run_predict.py --bias_name $BIAS_NAME --all_models $MODELS

Results analysis

Run the analysis script to create a .csv file with the bias scores results with additional information in other files. The results will be saved in The respected predictions dirs.

* Note that for the decoy analysis, you'll need to set BIAS_NAME=decoy_expensive or BIAS_NAME=decoy_cheaper for the respected biases analysis.

python run_analysis.py --bias_name $BIAS_NAME --all_models $MODELS

Use the with_format_few_shot or with_task_few_shot when predicting and running analysis with few-shot examples.

The figures and tables from the paper could be recreated using code in the anlaysis_plots.ipynb Jupyter notebook.

License

Instructed to Bias is MIT-licensed.