March 2025 Summaries

Predibase has launched the first end-to-end platform for reinforcement fine-tuning (RFT), aiming to make advanced model customization accessible to developers and enterprises by overcoming the common obstacle of limited labeled data. Reinforcement fine-tuning allows language models to learn from reward functions, optimizing performance for reasoning tasks and scenarios like code generation and complex reasoning, where traditional supervised fine-tuning falls short. The platform offers a fully-managed, serverless infrastructure that integrates the complete workflow from data to deployment, utilizing techniques such as supervised fine-tuning warm-ups, GRPO, and curriculum learning to enhance model performance. A notable achievement of this platform is its capacity to create specialized models, such as one that significantly outperformed larger models like OpenAI o1 and DeepSeek-R1 in a PyTorch-to-Triton code translation task, all while using fewer resources. The launch includes open-sourcing of the model on Hugging Face and invites developers to explore the platform's capabilities through demos and a webinar.

Large Language Models (LLMs) offer significant potential for enterprise efficiency, productivity, and savings, but fine-tuning is essential for specific use cases, as demonstrated by Marsh McLennan's implementation of LenAI. By working with Predibase to fine-tune their models, Marsh McLennan improved the accuracy and responsiveness of LenAI, an AI assistant designed to leverage institutional knowledge and provide industry expertise. This customization addressed challenges like intent recognition, which was problematic with off-the-shelf models such as GPT-3.5, and resulted in a 7-12% increase in accuracy and reduced latency. Under the leadership of Chief Information and Operations Officer Paul Beswick, the company adopted a proactive approach to generative AI, deploying APIs and launching LenAI to its global workforce, which now handles around 20 million requests annually. The success of LenAI has translated into a productivity boost, saving over 1 million hours in its first year, and has empowered employees to experiment with the technology to innovate further. By democratizing access to AI tools and fostering a culture of innovation, Marsh McLennan continues to expand LenAI’s capabilities and transform enterprise knowledge management.