New Hugging Face Models on Azure AI: Phi-3 Variants from the Community

Building generative AI applications starts with: Select model Choose the right model for your application requirements. that Azure AI Model Catalog It offers over 1,78,000 models, including foundational models from key partners and approximately 1,600 open source models from the Hugging Face community. This post is part of a monthly series aimed at raising awareness of new models added to Azure’s Hugging Face collection. Check out our previous content. hug face model recruitment post here.

that hug face model hub We have over 1 million models. Select up to 20 models to add to Azure each month. Based on feedback from our customers and developer community. Want to request that we add a specific Hugging Face model? Request in just 3 steps:

1. search hugging face hub For the model you want – click to see the model card.
2. On that page, click the “Deployment” dropdown and Azure ML Options.
3. In the pop-up dialog, look for the ‘Request more’ button to complete the flow.

If you already have a model in the Azure AI Model Catalog, you’ll see a “Go to model in Azure ML” button that takes you to the model card in Azure AI Studio.

18 new hug face models were added in September.

In September, 18 new models were added to Azure AI’s Hugging Face collection. Models included community-generated variants of popular base models, including Meta’s Llama family (LLM) and Microsoft’s Phi-3 family (SLM). We also note that many of these fine-tuned models “Made by Unsloth” – We’ll talk about this in a moment. Let’s first review the added models and highlight notable features.

Note: These community-created models are suitable for research and prototyping, but require further evaluation for use in production. Please read the model cards (linked to each) for usage instructions and limitations and perform your own quality and safety assessment to evaluate your specific application scenario.

Observed Topics

The 18 added models also help identify useful themes or trends in community-created variations in terms of use cases, tools, and processes. This is what we observed:

1. Multilingual models continue to shine – We’ve added models tailored for Spanish, Japanese, Korean, and Southeast Asian languages, many of which score well on our evaluation leaderboards. This highlights the growing demand for conversational tasks that can effectively reflect local vocabulary and culture.
2. Phi-3 variants continue to grow – Microsoft’s Phi-3 family of “small language models” (SLMs) outperforms other similar models in the same or adjacent size class. More variants are now being developed, potentially for scenarios in mobile and edge devices. Find out more about this below.
3. Fine-tuning tools have value – Community-written variants focus on fine-tuning popular base models, but this is time-intensive and expensive. We are now seeing more models. laggard” reflects interest in tools and processes that speed up fine-tuning with less memory without sacrificing accuracy. Find out more about this below.
4. Meta/Rama are still popular – The Meta/Llama family of models continues to influence community creators in a variety of ways. The first is a base model for fine-tuning (e.g. multilingual from SEA-lion, using tools from Groq) and the second is a target for optimization (e.g. TinyLlama suite for mobile and edge devices). We can also see adapting other base models (e.g. the Llamafed version of Phi-3) to adapt that model to familiar structures.

Model Spotlight: Phi-3 Community Variants

that hug face model hub Phi-3 variants are created every day. In general, the Phi-3 model family outperforms other models in its size class (and adjacent ones), making it ideal for use cases targeting edge and mobile devices. This month we added seven models to Azure, fine-tuned but with different goals: the Phi-3 and Phi-3.5 base models. Let us briefly review these.

The first variant recalibrates the model to the Llama2/Llama3 model structure for developer friendliness. The second uses AutoAWQ with 4-bit quantization to obtain a model that can operate on smaller GPUs. The third uses examples generated from Llama-3 for fine-tuning and scores well on two popular benchmarks. Variants 4 and 7 are from Unsloth and demonstrate the fine-tuning techniques described in the next section. Variant 5 shows an example of fine-tuning Phi-3 to “supplement” and not censor information. Variant 6 demonstrates the math capabilities of Phi-3 fine-tuned with the popular Orca Math Word problem dataset.

want to explore Want to explore Phi-3 features but don’t know where to start? Add to bookmarks Pi-3 Cookbook Starting from Microsoft Welcome to the Phi-3 family page. Then index Links to quickstarts, tutorials, and E2E samples. Then try one of the fine-tuned variations above to see the difference.

A call to action for the community

1. Help us shine a light on your work!

Have you built interesting AI applications using the Hugging Face model on Azure? Have you ever published your own fine-tuned variations of a popular basic model? We’d like to know more. Please leave a comment on this blog with a link to your article or repository. We want to learn more and expand our community of model makers.

2. Getting started with the hug face model in Azure

Are you new to the Azure AI model catalog and want to start using the Hugging Face model in Azure? Here are three resources to help you start your learning journey:

1. Azure AI Model Catalog: Explore the Hugging Face collection
2. Azure ML documentation: Model Catalogs and Collections
3. Azure ML sample notebooks: Explore inference tasks using code

Note: We’re adding up to 20 models to the Hugging Face collection on Azure each month, and we want your feedback to help us make these decisions! Request a model using the three-step process outlined earlier in the article and tell us more about how you’re currently using it!