Announcing Phi-3 fine-tuning, new generative AI models, and other Azure AI updates to empower organizations to customize and scale AI applications

5 months ago 25
News Banner

Looking for an Interim or Fractional CTO to support your business?

Read more

AI is transforming each manufacture and creating caller opportunities for innovation and growth. But, processing and deploying AI applications astatine standard requires a robust and flexible level that tin grip the analyzable and divers needs of modern enterprises and let them to make solutions grounded successful their organizational data. That’s wherefore we are excited to denote respective updates to assistance developers rapidly make customized AI solutions with greater prime and flexibility leveraging the Azure AI toolchain:

  • Serverless fine-tuning for Phi-3-mini and Phi-3-medium models enables developers to rapidly and easy customize the models for unreality and borderline scenarios without having to put for compute.
  • Updates to Phi-3-mini including significant betterment successful halfway quality, instruction-following, and structured output, enabling developers to physique with a much performant exemplary without further cost.
  • Same time shipping earlier this period of the latest models from OpenAI (GPT-4o mini), Meta (Llama 3.1 405B), Mistral (Large 2) to Azure AI to supply customers greater prime and flexibility.

Unlocking worth done exemplary innovation and customization  

In April, we introduced the Phi-3 family of small, unfastened models developed by Microsoft. Phi-3 models are our astir susceptible and cost-effective tiny connection models (SLMs) available, outperforming models of the aforesaid size and adjacent size up. As developers look to tailor AI solutions to conscionable circumstantial concern needs and amended prime of responses, fine-tuning a tiny exemplary is simply a large alternate without sacrificing performance. Starting today, developers tin fine-tune Phi-3-mini and Phi-3-medium with their information to physique AI experiences that are much applicable to their users, safely, and economically.

Given their tiny compute footprint, unreality and borderline compatibility, Phi-3 models are good suited for fine-tuning to amended basal exemplary show crossed a assortment of scenarios including learning a caller accomplishment oregon a task (e.g. tutoring) oregon enhancing consistency and prime of the effect (e.g. code oregon benignant of responses successful chat/Q&A). We’re already seeing adaptations of Phi-3 for caller usage cases.

Blue cylinder with purple background

Phi-3 models

A household of powerful, tiny connection models (SLMs) with groundbreaking show astatine debased outgo and debased latency

Microsoft and Khan Academy are moving unneurotic to assistance amended solutions for teachers and students crossed the globe. As portion of the collaboration, Khan Academy uses Azure OpenAI Service to power Khanmigo for Teachers, a aviator AI-powered teaching adjunct for educators crossed 44 countries and is experimenting with Phi-3 to amended mathematics tutoring. Khan Academy precocious published a probe insubstantial highlighting however antithetic AI models execute erstwhile evaluating mathematical accuracy successful tutoring scenarios, including benchmarks from a fine-tuned mentation of Phi-3. Initial data shows that erstwhile a pupil makes a mathematical error, Phi-3 outperformed astir different starring generative AI models astatine correcting and identifying pupil mistakes.

And we’ve fine-tuned Phi-3 for the instrumentality too. In June, we introduced Phi Silica to empower developers with a powerful, trustworthy exemplary for gathering apps with safe, unafraid AI experiences. Phi Silica builds connected the Phi household of models and is designed specifically for the NPUs successful Copilot+ PCs. Microsoft Windows is the archetypal level to person a state-of-the-art tiny connection exemplary (SLM) customized built for the Neural Processing Unit (NPU) and shipping inbox.

You tin effort fine-tuning for Phi-3 models contiguous successful Azure AI.

I americium besides excited to stock that our Models-as-a-Service (serverless endpoint) capableness successful Azure AI is present mostly available. Additionally, Phi-3-small is present disposable via a serverless endpoint truthful developers tin rapidly and easy get started with AI improvement without having to negociate underlying infrastructure. Phi-3-vision, the multi-modal exemplary successful the Phi-3 family, was announced astatine Microsoft Build and is disposable done Azure AI exemplary catalog. It volition soon beryllium disposable via a serverless endpoint arsenic well. Phi-3-small (7B parameter) is disposable successful 2 discourse lengths 128K and 8K whereas Phi-3-vision (4.2B parameter) has besides been optimized for illustration and diagram knowing and tin beryllium utilized to make insights and reply questions.

We are seeing large effect from the assemblage connected Phi-3. We released an update for Phi-3-mini past period that brings important betterment successful halfway prime and acquisition following. The exemplary was re-trained starring to important betterment successful acquisition pursuing and enactment for structured output. We besides improved multi-turn speech quality, introduced enactment for <|system|> prompts, and importantly improved reasoning capability.

The array beneath highlights improvements crossed acquisition following, structured output, and reasoning.

Benchmarks Phi-3-mini-4k Phi-3-mini-128k 
Apr ’24 release Jun ’24 update Apr ’24 release Jun ’24 update 
Instruction Extra Hard 5.7 6.0 5.7 5.9 
Instruction Hard 4.9 5.1 5.2 
JSON Structure Output 11.5 52.3 1.9 60.1 
XML Structure Output 14.4 49.8 47.8 52.9 
GPQA 23.7 30.6 25.9 29.7 
MMLU 68.8 70.9 68.1 69.7 
Average 21.7 35.8 25.7 37.6 

We proceed to marque improvements to Phi-3 information too. A caller probe paper highlighted Microsoft’s iterative “break-fix” attack to improving the information of the Phi-3 models which progressive aggregate rounds of investigating and refinement, reddish teaming, and vulnerability identification. This method importantly reduced harmful contented by 75% and enhanced the models’ show connected liable AI benchmarks. 

Expanding exemplary choice, present with implicit 1600 models disposable successful Azure AI

With Azure AI, we’re committed to bringing the astir broad enactment of unfastened and frontier models and state-of-the-art tooling to assistance conscionable customers’ unsocial cost, latency, and plan needs. Last twelvemonth we launched the Azure AI exemplary catalog wherever we present person the broadest enactment of models with implicit 1,600 models from providers including AI21, Cohere, Databricks, Hugging Face, Meta, Mistral, Microsoft Research, OpenAI, Snowflake, Stability AI and others. This period we added—OpenAI’s GPT-4o mini done Azure OpenAI Service, Meta Llama 3.1 405B, and Mistral Large 2.

Continuing the momentum contiguous we are excited to stock that Cohere Rerank is present disposable connected Azure. Accessing Cohere’s enterprise-ready connection models connected Azure AI’s robust infrastructure enables businesses to seamlessly, reliably, and safely incorporated cutting-edge semantic hunt exertion into their applications. This integration allows users to leverage the flexibility and scalability of Azure, combined with Cohere’s highly performant and businesslike connection models, to present superior hunt results successful production.

TD Bank Group, 1 of the largest banks successful North America, precocious signed an statement with Cohere to research its afloat suite of ample connection models (LLMs), including Cohere Rerank.

At TD, we’ve seen the transformative imaginable of AI to present much personalized and intuitive experiences for our customers, colleagues and communities, we’re excited to beryllium moving alongside Cohere to research however its connection models execute connected Microsoft Azure to assistance enactment our innovation travel astatine the Bank.”

Kirsti Racine, VP, AI Technology Lead, TD.

Atomicwork, a integer workplace acquisition level and longtime Azure customer, has importantly enhanced its IT work absorption level with Cohere Rerank. By integrating the exemplary into their AI integer assistant, Atom AI, Atomicwork has improved hunt accuracy and relevance, providing faster, much precise answers to analyzable IT enactment queries. This integration has streamlined IT operations and boosted productivity crossed the enterprise. 

The driving unit down Atomicwork’s integer workplace acquisition solution is Cohere’s Rerank exemplary and Azure AI Studio, which empowers Atom AI, our integer assistant, with the precision and show required to present real-world results. This strategical collaboration underscores our committedness to providing businesses with advanced, secure, and reliable endeavor AI capabilities.”

Vijay Rayapati, CEO of Atomicwork

Command R+, Cohere’s flagship generative exemplary which is besides disposable connected Azure AI, is purpose-built to enactment good with Cohere Rerank wrong a Retrieval Augmented Generation (RAG) system. Together they are susceptible of serving immoderate of the astir demanding endeavor workloads successful production. 

Earlier this week, we announced that Meta Llama 3.1 405B on with the latest fine-tuned Llama 3.1 models, including 8B and 70B, are present disposable via a serverless endpoint successful Azure AI. Llama 3.1 405B tin beryllium utilized for precocious synthetic information procreation and distillation, with 405B-Instruct serving arsenic a teacher exemplary and 8B-Instruct/70B-Instruct models acting arsenic pupil models. Learn much astir this announcement here.

Mistral Large 2 is present disposable connected Azure, making Azure the archetypal starring unreality supplier to connection this next-gen model. Mistral Large 2 outperforms erstwhile versions successful coding, reasoning, and agentic behavior, lasting connected par with different starring models. Additionally, Mistral Nemo, developed successful collaboration with NVIDIA, brings a almighty 12B exemplary that pushes the boundaries of connection knowing and generation. Learn More.

And past week, we brought GPT-4o mini to Azure AI alongside different updates to Azure OpenAI Service, enabling customers to grow their scope of AI applications astatine a little outgo and latency with improved information and information deployment options. We volition denote much capabilities for GPT-4o mini successful coming weeks. We are besides blessed to present a caller diagnostic to deploy chatbots built with Azure OpenAI Service into Microsoft Teams.  

Enabling AI innovation safely and responsibly  

Building AI solutions responsibly is astatine the halfway of AI improvement astatine Microsoft. We person a robust acceptable of capabilities to assistance organizations measure, mitigate, and negociate AI risks crossed the AI improvement lifecycle for accepted instrumentality learning and generative AI applications. Azure AI evaluations alteration developers to iteratively measure the prime and information of models and applications utilizing built-in and customized metrics to pass mitigations. Additional Azure AI Content Safety features—including punctual shields and protected worldly detection—are present “on by default” successful Azure OpenAI Service. These capabilities tin beryllium leveraged arsenic contented filters with immoderate instauration exemplary included successful our exemplary catalog, including Phi-3, Llama, and Mistral. Developers tin besides integrate these capabilities into their exertion easy done a azygous API. Once successful production, developers tin monitor their application for prime and safety, adversarial punctual attacks, and information integrity, making timely interventions with the assistance of real-time alerts.

Azure AI uses HiddenLayer Model Scanner to scan third-party and unfastened models for emerging threats, specified arsenic cybersecurity vulnerabilities, malware, and different signs of tampering, earlier onboarding them to the Azure AI exemplary catalog. The resulting verifications from Model Scanner, provided wrong each exemplary card, tin springiness developer teams greater assurance arsenic they select, fine-tune, and deploy unfastened models for their application. 

We proceed to put crossed the Azure AI stack to bring authorities of the creation innovation to our customers truthful you tin build, deploy, and standard your AI solutions safely and confidently. We cannot hold to spot what you physique next.

Stay up to day with much Azure AI news

  • Watch this video to larn much astir Azure AI exemplary catalog.
  • Listen to the podcast connected Phi-3 with pb Microsoft researcher Sebastien Bubeck.

The station Announcing Phi-3 fine-tuning, caller generative AI models, and different Azure AI updates to empower organizations to customize and standard AI applications appeared archetypal connected Microsoft Azure Blog.

Read Entire Article