Hugging Face Private Hub Introduction

Hugging Face has launched the Private Hub (now rebranded as the Enterprise Hub), providing a secure and compliant ecosystem for organizations to build, collaborate, and deploy machine learning (ML) models. The platform aims to solve common industry bottlenecks—such as duplicated efforts, non-standard workflows, and the difficulty of moving models from research to production—by unifying the ML lifecycle in one place.

The Private Hub Ecosystem

The Private Hub allows companies to utilize the full Hugging Face ecosystem within a private environment. This ensures that sensitive data and models remain on the company's own servers while leveraging the following core components:

Core Assets

  • Models: Access to state-of-the-art pre-trained models across various domains (NLP, computer vision, speech, etc.) that can be fine-tuned for specific use cases via transfer learning.
  • Datasets: Hosting for proprietary datasets with integrated tools for efficient data manipulation and viewing.
  • Spaces: A framework for building and hosting interactive ML demo apps to showcase work to stakeholders and gather early feedback.

Integrated Tools

  • AutoTrain: A no-code AutoML solution for automatically fine-tuning, evaluating, and deploying models.
  • Evaluate: A tool to evaluate models on any dataset using various metrics without writing code.
  • Inference API: A system that deploys private models for inference via simple API calls, removing the need for manual Docker or Kubernetes management.
  • Collaboration Tools: Git-based repositories with support for pull requests and discussions to facilitate peer reviews and version control.

Deployment Options

To meet varying security and compliance requirements, Hugging Face provides three deployment models for the Private Hub:

  1. Managed Private Hub (SaaS): Hosted in segregated virtual private servers (VPCs) owned by Hugging Face, removing infrastructure management overhead.
  2. On-cloud Private Hub: Deployed within a customer's own cloud account (AWS, Azure, or GCP), providing full administrative control over the infrastructure.
  3. On-prem Private Hub: Deployed on the customer's own physical infrastructure for those with the strictest compliance rules. Note: As of June 2023, on-prem deployments were discontinued as part of the transition to the Enterprise Hub.

Accelerating the ML Roadmap: A Practical Workflow

Hugging Face illustrates the Private Hub's utility through a financial analyst report analysis use case, demonstrating a streamlined transition from research to production:

1. Rapid Model Development

Instead of training from scratch, teams can clone a pre-trained model (e.g., FinBERT) into their Private Hub and fine-tune it using a custom, proprietary dataset. By using AutoTrain, teams can automatically train and compare multiple candidate models (e.g., 25 different versions) to identify the best performer based on metrics and the Inference Widget.

2. Stakeholder Validation

Using Spaces, teams can build interactive demo apps in minutes (e.g., using approximately 51 lines of code) to validate a Minimum Viable Product (MVP) with business stakeholders, enabling faster iteration based on real-world feedback.

3. Production Scaling

Once a model is approved by compliance teams via model cards, it can be deployed to production using the Inference API. This allows teams to run inferences at scale and in real-time using simple HTTP requests (approximately 12 lines of code), bypassing the complexities of MLOps infrastructure management.

Transition to Enterprise Hub

As of June 2023, the Private Hub has been renamed the Enterprise Hub. This hosted solution combines SaaS convenience with enterprise security, offering advanced user administration, SSO, and the ability to deploy services like Inference Endpoints across various compute options, including on-cloud and on-prem.

Sources