OpenAI's Reinforcement Fine-Tuning Program Boosts o1-mini's Accuracy by 31%

Dec 6, 2024, 06:22 PM

OpenAI has announced the launch of a new Reinforcement Fine-Tuning (RFT) program, allowing developers to customize AI models for specific domain tasks with minimal training data. This initiative enables the creation of expert models in fields like legal, finance, engineering, biology, and insurance using only a few dozen examples. The RFT process trains models to reason in new ways over custom domains, rather than merely memorizing answers. OpenAI demonstrated the capabilities of this approach with the o1-mini model, which, after fine-tuning, outperformed both its base version and the standard o1 model in tasks such as rare disease gene prediction, achieving a 31% higher accuracy rate. This development is part of a broader expansion of OpenAI's Reinforcement Fine-Tuning Research Program, aimed at enhancing model performance in specialized areas.