---
license: mit
---

This repository hosts the under-trained [detoxify](https://github.com/unitaryai/detoxify) models used as value models in the experiments in the paper [Language Model Decoding as Likelihood-Utility Alignment](https://arxiv.org/abs/2210.07228) testing the robustness to noise of value-guided decoding algorithms (MCTS and VGBS). 
For more details see the paper or the [project's homepage and GitHub repository](https://github.com/epfl-dlab/understanding-decoding).