HelpSteer2-Preference: Complementing Ratings with Preferences

• URL: http://arxiv.org/abs/2410.01257v1
• Date: Wed, 2 Oct 2024 06:05:52 GMT
• Title: HelpSteer2-Preference: Complementing Ratings with Preferences
• Authors: Zhilin Wang, Alexander Bukharin, Olivier Delalleau, Daniel Egert, Gerald Shen, Jiaqi Zeng, Oleksii Kuchaiev, Yi Dong,
• Abstract summary: Reward models are critical for aligning models to follow instructions. There is a lack of evidence that either approach is better than the other when adequately matched for data. We propose a novel approach to combine Bradley-hugging and Regression reward modeling.
• Score: 45.01567242039055
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. We also demonstrate the effectiveness of this reward model at aligning models to follow instructions in RLHF. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2 and openly release the trained Reward Model at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward

Related papers

• Optimal Design for Reward Modeling in RLHF [83.3614658277817]
  We formalize the reward training model in Reinforcement Learning from Human Feedback. We frame the selection of an effective dataset as a simple regret minimization task. We derive bounds on the simple regret under appropriate assumptions.
  arXiv  Detail & Related papers  (2024-10-22T14:36:44Z)
• General Preference Modeling with Preference Representations for Aligning Language Models [51.14207112118503]
  We introduce preference representation learning, an approach that embeds responses into a latent space to capture intricate preference structures efficiently. We also propose preference score-based General Preference Optimization (GPO), which generalizes reward-based reinforcement learning from human feedback. Our method may enhance the alignment of foundation models with nuanced human values.
  arXiv  Detail & Related papers  (2024-10-03T04:22:55Z)
• Preference Alignment with Flow Matching [23.042382086241364]
  Preference Flow Matching (PFM) is a new framework for preference-based reinforcement learning (PbRL) It streamlines the integration of preferences into an arbitrary class of pre-trained models. We provide theoretical insights that support our method's alignment with standard PbRL objectives.
  arXiv  Detail & Related papers  (2024-05-30T08:16:22Z)
• RewardBench: Evaluating Reward Models for Language Modeling [100.28366840977966]
  We present RewardBench, a benchmark dataset and code-base for evaluation of reward models. The dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods.
  arXiv  Detail & Related papers  (2024-03-20T17:49:54Z)
• Differentially Private Reward Estimation with Preference Feedback [15.943664678210146]
  Learning from preference-based feedback has recently gained considerable traction as a promising approach to align generative models with human interests. An adversarial attack in any step of the above pipeline might reveal private and sensitive information of human labelers. We focus on the problem of reward estimation from preference-based feedback while protecting privacy of each individual labelers.
  arXiv  Detail & Related papers  (2023-10-30T16:58:30Z)
• Zephyr: Direct Distillation of LM Alignment [59.03530095974505]
  We aim to produce a smaller language model that is aligned to user intent. Previous research has shown that applying supervised fine-tuning (dSFT) on larger models significantly improves task accuracy. We apply distilled direct preference optimization (dDPO) to learn a chat model with significantly improved intent alignment.
  arXiv  Detail & Related papers  (2023-10-25T19:25:16Z)
• Backward Compatibility During Data Updates by Weight Interpolation [17.502410289568587]
  We study the problem of regression during data updates and propose Backward Compatible Weight Interpolation (BCWI) BCWI reduces negative flips without sacrificing the improved accuracy of the new model. We also explore the use of importance weighting during and averaging the weights of multiple new models in order to further reduce negative flips.
  arXiv  Detail & Related papers  (2023-01-25T12:23:10Z)
• Variational Bayesian Unlearning [54.26984662139516]
  We study the problem of approximately unlearning a Bayesian model from a small subset of the training data to be erased. We show that it is equivalent to minimizing an evidence upper bound which trades off between fully unlearning from erased data vs. not entirely forgetting the posterior belief. In model training with VI, only an approximate (instead of exact) posterior belief given the full data can be obtained, which makes unlearning even more challenging.
  arXiv  Detail & Related papers  (2020-10-24T11:53:00Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.