OpenAI has developed CriticGPT, a new model designed to identify errors in code generated by ChatGPT, aiming to improve the accuracy of large language models (LLMs). Typically, Reinforcement Learning from Human Feedback (RLHF) is used to refine the output, but this process can be time-consuming and error-prone, especially with large models. CriticGPT, based on GPT-4, has been shown to outperform humans in reviewing ChatGPT code, detecting both common and less frequent bugs.

The model has been trained on a dataset of code samples with intentional bugs and feedback, enabling it to detect errors more effectively than the average human code reviewer. CriticGPT generates fewer false positives and nitpicks less about code compared to humans, making it a valuable tool for improving code accuracy.

OpenAI plans to integrate CriticGPT-like models into its RLHF labeling pipeline to assist model trainers. While the results presented are from a research phase, the potential for CriticGPT to enhance the accuracy of LLMs is promising. The model’s ability to detect errors and provide valuable feedback makes it a valuable tool for improving the performance of language models.

Source link

Source link: https://www.techzine.eu/news/analytics/121740/criticgpt-finds-errors-in-the-output-of-ai-models/