Microsoft has up to date its Azure AI Foundry portal and Azure OpenAI Service APIs and SDKs to help Direct Choice Optimization (DPO) for GPT-4.1 and GPT-4.1-mini. Direct Choice Optimization (DPO) is a fine-tuning method that can be utilized to regulate mannequin weights primarily based on human preferences utilizing a pair of most well-liked and non-preferred responses.

One of many major advantages of utilizing DPO over Reinforcement Studying from Human Suggestions (RLHF) is that it’s computationally lighter and quicker whereas being simply as efficient for mannequin alignment. Organizations can use this technique to coach fashions to match their particular model voice, security necessities, or conversational kinds.

Along with utilizing DPO for mannequin fine-tuning, Microsoft has expanded Azure AI’s World Coaching to 12 new areas together with East US, West Europe, UK South, Switzerland North, and extra. Regardless of the growth, it’s nonetheless thought-about to be a public preview.

Microsoft stated that customers ought to anticipate and watch for brand new options coming quickly together with pause/resume performance and steady fine-tuning. It can even be bringing GPT-4.1-nano to those new areas.

The growth of World Coaching is vital for knowledge sovereignty, which is changing into a extra vital challenge with the European Union pushing for European’s knowledge to be dealt with in Europe to make sure higher privateness.

Lastly, Microsoft has launched the brand new Responses API which helps your fine-tuned fashions, making it simpler for builders to make use of them within different purposes. Microsoft stated this API is good for agentic workflows as “it helps stateful, multi-turn conversations and permits seamless instrument calling, mechanically stitching every little thing collectively within the background.”

The Responses API also can hold monitor of conversations in order that the mannequin can keep in mind context, you’ll be able to see how fashions motive by means of solutions, it will possibly let customers test the progress whereas a response generates, and it helps background processing and works with instruments like net search and file lookup.

Picture by way of Depositphotos.com