In the field of Artificial Intelligence (AI), accurate assessment of model outputs is crucial. OpenAI researchers have developed CriticGPT, a tool that helps human trainers identify errors in AI-generated responses, particularly in code outputs. CriticGPT aims to improve the precision and reliability of AI systems by providing thorough criticisms. In experiments, human reviewers using CriticGPT performed significantly better than those without it. This tool is being integrated into the Reinforcement Learning with Human Feedback (RLHF) labeling pipeline to enhance the evaluation of advanced AI system outputs.
ChatGPT, powered by the GPT-4 series, relies on human trainers to evaluate responses and gather comparative data. As AI models become more complex, identifying errors becomes increasingly challenging. CriticGPT assists in pinpointing errors in ChatGPT’s responses, ultimately improving the assessment process and ensuring that AI models align with their intended behaviors and goals. The team has made significant contributions by offering a scalable oversight technique that helps humans detect problems in RLHF data more thoroughly. CriticGPT has proven to be effective in catching errors and producing more thorough criticisms compared to human contractors working alone. The integration of CriticGPT into the training process aims to enhance the evaluation of AI outputs and improve human-AI cooperation.
Source link
Source link: https://www.marktechpost.com/2024/06/28/openai-introduces-criticgpt-a-new-artificial-intelligence-ai-model-based-on-gpt-4-to-catch-errors-in-chatgpts-code-output/?amp