Official Code For the Paper "More RLHF, More Trust? On The Impact of Human Preference Alignment On Language Model Trustworthiness"

Overview

This repo contains the experiment code for the paper "More RLHF, More Trust? On The Impact of Human Preference Alignment On Language Model Trustworthiness". See our paper for details.

This project is organized around the following five primary areas of trustworthiness, including:

1. Toxicity
2. Stereotypical bias
3. Machine Ethics
4. Truthfulness
5. Privacy

Project Structure

This project is structured around subdirectories dedicated to each area of trustworthiness. Each subdir includes python scripts to reproduce our experiment results. Additionally, the data for each evaluation is included in the data/ subdirectory.

.
├── README.md
├── data/
    ├── toxicity/
    ├── stereotype/
    ├── machine_ethics/
    ├── truthfulness/
    ├── privacy/
├── toxicity_eval/
├── stereotype_eval/
├── machine_ethics/
├── privacy/
├── machine_ethics/

(⚠️WARNING: This paper contains model outputs that may be considered offensive.)

In general, python scripts end with _exp.py are used for language model inference, and _analysis.py are used for calculating the numerical benchmark values. Other area-specific analysis will be explained in the corresponding subdirectory.

[Important] Candidate models

In this project, we mainly focus on the open source Pythia suite, and here are the links to the specific models hosted on Huggingface

• Pythia 70M: EleutherAI/pythia-70m
• Pythia 160M: EleutherAI/pythia-160m
• Pythia 410M: EleutherAI/pythia-410m
• Pythia 2.8B: EleutherAI/pythia-2.8b
• Pythia 6.9B: EleutherAI/pythia-6.9b