https://github.com/RLHF-V/RLAIF-V

[CVPR'25] RLAIF-V: Aligning MLLMs through Open-Source AI Feedback for Super GPT-4V Trustworthiness
https://github.com/RLHF-V/RLAIF-V

chatbot cvpr2025 gpt-4v llava llava-next minicpm-v multimodal rlaif-v vision-language-learning

Last synced: about 1 month ago
JSON representation

[CVPR'25] RLAIF-V: Aligning MLLMs through Open-Source AI Feedback for Super GPT-4V Trustworthiness

• Host: GitHub
• URL: https://github.com/RLHF-V/RLAIF-V
• Owner: RLHF-V
• Created: 2024-05-13T07:58:37.000Z (12 months ago)
• Default Branch: main
• Last Pushed: 2025-03-03T11:40:28.000Z (2 months ago)
• Last Synced: 2025-03-03T12:34:19.562Z (2 months ago)
• Topics: chatbot, cvpr2025, gpt-4v, llava, llava-next, minicpm-v, multimodal, rlaif-v, vision-language-learning
• Language: Python
• Homepage:
• Size: 56.4 MB
• Stars: 300
• Watchers: 5
• Forks: 12
• Open Issues: 7
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• awesome-RLAIF - Datasets and models
• StarryDivineSky - RLHF-V/RLAIF-V - V是一个开源项目，旨在通过人工智能反馈（AI Feedback）提升GPT-4V等视觉语言模型的可靠性和安全性。该项目基于RLAIF（Reinforcement Learning from AI Feedback）框架，利用AI而非人类来评估和改进模型的行为。其核心思想是训练一个奖励模型，该模型能够判断模型输出的好坏，并用此奖励信号来优化视觉语言模型。项目特色在于其开源性和对GPT-4V等先进模型的适配，目标是使AI系统更加可信赖。具体实现包括数据收集、奖励模型训练和强化学习优化三个阶段。项目代码和预训练模型将会开源，方便研究人员复现和进一步研究。该项目是CVPR 2025的论文成果，表明其在计算机视觉领域的学术价值。通过使用AI反馈，RLAIF-V有望减少人工干预，并提升模型训练的效率和可扩展性。 (多模态大模型 / 资源传输下载)