README.md

Aligning Language Models with Human Preferences via a Bayesian Approach

This is the PyTorch implementation of the paper:

Aligning Language Models with Human Preferences via a Bayesian Approach.

Jiashuo WANG, Haozhao WANG, Shichao SUN, Wenjie LI, NeurIPS 2023

If you use our codes or your research is related to our work, please kindly cite our paper:

@inproceedings{wang-etal-2023-aligning,
    title = "Aligning Language Models with Human Preferences via a Bayesian Approach",
    author = "Wang, Jiashuo  and
      Wang, Haozhao  and
      Sun, Shichao and
      Li, Wenjie",
    booktitle = "Conference on Neural Information Processing Systems: NeurIPS 2023",
    month = dec,
    year = "2023",
    address = "New Orleans, LA, USA",
    organization= "PMLR"
}

Abstract

In the quest to advance human-centric natural language generation (NLG) systems, ensuring alignment between NLG models and human preferences is crucial. For this alignment, current popular methods leverage a reinforcement learning (RL) approach with a reward model trained on feedback from humans. However, inherent disagreements due to the subjective nature of human preferences pose a significant challenge for training the reward model, resulting in a deterioration of the NLG performance. To tackle this issue, previous approaches typically rely on majority voting or averaging to consolidate multiple inconsistent preferences into a merged one. Although straightforward to understand and execute, such methods suffer from an inability to capture the nuanced degrees of disaggregation among humans and may only represent a specialized subset of individuals, thereby lacking the ability to quantitatively disclose the universality of human preferences. To address this challenge, this paper proposes a novel approach, which employs a Bayesian framework to account for the distribution of disagreements among human preferences as training a preference model, and names it as d-PM. Besides, considering the RL strategy's inefficient and complex training process over the training efficiency, we further propose utilizing the contrastive learning strategy to train the NLG model with the preference scores derived from the d-PM model. Extensive experiments on two human-centric NLG tasks, i.e., emotional support conversation and integrity "Rule-of-Thumb" generation, show that our method consistently exceeds previous SOTA models in both automatic and human evaluations.

Model Architecture:

Preparing Environment

conda env create -f env.yml -n alignment
conda activate alignment

Experiments

Preference Modeling with Disagreement

Refer to the preference_modeling folder for more details.

Calibration for Alignment

Emotional Support Conversation: Refer to the ESConv and MultiESC folders.

Integrity RoT Generation: Refer to the mic folder.

Checkpoints

The checkpoints can be downloaded from here.