We are particularly grateful to Daniel Benito, CTO at Bitext, Antonio Valderrabanos, CEO at Bitext, Chen Wang, Lead Resolution Architect at AI21 Labs, Robbin Jang, Alliance Supervisor at AI21 Labs, and Alex Godfrey, Companion Advertising Lead at AI21 Labs, for their valuable insights and contributions to this blog.
The common availability of AI mannequin sharing is now accessible within Databricks Delta Sharing and the Databricks Marketplace. This milestone follows the . To mark the successful Public Preview launch, we’ve collaborated with innovative AI model-sharing platforms like TensorFlow, Hugging Face, and Ripple to further streamline AI Model Sharing.
You can easily share and serve AI models securely using Delta Sharing. Sharing can occur within your group or extend externally across clouds, platforms, and regions. As well as, the platform now boasts over 75+ AI models, including industry-specific AI models from John Snow Labs, OLA Krutrim, Bitext, in addition to foundational models like Databricks DBRX, Llama-3, AI21 Labs, Mistral, and several others.
This blog post will delve into the business imperative driving the need for AI model sharing, with a focus on practical applications powered by AI21’s Jamba 1.5 Mini foundation model and Bitext models.
AI-powered fashion solutions are now readily accessible out of the box from our platform, significantly simplifying the process for customers to access and deploy models seamlessly. As this advancement unfolds, it not only streamlines user experience but also amplifies the availability of AI models, facilitating effortless interconnection and deployment across diverse platforms and geographies?
Three key benefits of AI Model Sharing with Databricks that emerged from our conversations with early adopters and launch partners include:
- By leveraging AI-powered mannequins in conjunction with Delta Sharing, asset owners can significantly reduce their total cost of ownership through minimized expenses associated with acquisition, growth, and infrastructure upkeep. Organisations can leverage pre-built or third-party solutions, including Delta Shared models or those sourced from Databricks Market, thereby decreasing initial investment and accelerating growth. By leveraging Delta Sharing across clouds and platforms, organizations can optimize their infrastructure utilization, reduce redundancy and costs, while simultaneously deploying fashions closer to end-users to minimize latency and improve overall performance.
- Delta Sharing empowers seamless integration of fashion offerings tailored to specific usage scenarios, streamlining the entire AI development process within a unified platform. By integrating their fashion models with the Databricks Mosaic AI platform, customers gain access to advanced AI and governance capabilities that enable the productionization of any model. The platform features comprehensive end-to-end model development capabilities, encompassing training to fine-tuning, as well as Unity Catalog’s robust security and management options such as lineage tracking and Lakehouse monitoring, ensuring exceptionally high confidence in models and associated data.
- When collaborating with external designers, AI-powered model-sharing solutions grant users complete control over associated fashion designs and data assets. Thanks to Delta Sharing, customers can assemble entire model datasets, allowing the model and data to reside within the buyer’s infrastructure, under their full control and oversight. Companies don’t need to transmit sensitive data to a vendor who is essentially acting as a proxy for their customer.
AI-Mannequin Sharing is enabled through our partnership with Delta Sharing technology. Suppliers can seamlessly share AI-driven fashion templates with potential clients using either Delta Sharing’s instant sharing capabilities or by listing them on the Databricks Marketplace, a platform that leverages Delta Sharing for seamless collaboration.
Delta Sharing empowers seamless access to AI models across any environment. One can apply fashion principles anywhere, subsequently utilizing them everywhere without needing to manually move them around. The mannequin weights (i.e. The parameters that the AI model has learned during training will automatically feed into the serving endpoint (i.e., The display window where the mannequin “lives”). This simplifies the need for complex manual adjustments following each model training or fine-tuning session, ensuring a unified source of truth and streamlining the serving process. Prospects can refine their fashion choices within the cloud, leveraging a cost-effective infrastructure for training AI models. Subsequently, they can deploy the model in another region closer to end-users, thereby minimizing inference latency – the time it takes for an AI model to process data and deliver results.
Databricks Marketplace, powered by Delta Sharing, simplifies the discovery and utilization of shared data assets. These styles can likely be rearranged to conform to your native system, with Delta Sharing automatically updating them during deployments or upgrades? You can also customize fashion styles together with your own information to fulfill duties such as managing a database. As a supplier, I seek to provide a single instance of my mannequin that can be shared seamlessly across all my Databricks customers.
As the Public Preview of AI Mannequin Sharing launched in January 2024, our team collaborated closely with numerous partners and customers to ensure that this innovative solution delivers significant cost savings and tangible value to organizations.
“We leverage Reinforcement Learning algorithms in select products to drive innovation.” Unlike traditional supervised learning approaches, reinforcement learning methods require significantly more extensive training periods and are inherently plagued by numerous sources of randomness throughout the training process. The RL fashions must be deployed across three distinct workspaces, situated in separate Amazon Web Services (AWS) regions. Through seamless mannequin sharing, users can access a single, fully trained RL model across multiple workspaces without requiring additional training or tedious manual processes to deploy it.
Mihir Mavalankar — A Machine Learning Engineer at Ripple
The AI21 Labs, a pioneer in the development of generative artificial intelligence and large language models, has announced its latest breakthroughs. AI21 Labs’ Jamba 1.5 Mini revolutionizes the application of AI-powered language models in business settings with its pioneering approach. The innovative hybrid Mamba-Transformer architecture enables a 256K-token efficient context window, characterized by its unique pace and exceptional quality. With its environmental optimization, Mini enables efficient computing by handling context lengths of up to 140K tokens on a single GPU.
“AI21 Labs is pleased to announce that Jamba 1.5 Mini is now available on the Databricks Marketplace.” Enterprises can seamlessly integrate their solutions with Delta Sharing, leveraging the advanced architecture of our Mamba-Transformer structure, which features a 256KB context window, enabling distinct speed and superior quality for transformative AI applications.
— Pankaj Dugar, SVP & GM , AI21 Labs
In AI architectures, a 256K token-efficient context window denotes the model’s ability to process and consider up to 256,000 tokens of textual content simultaneously. That is crucial as it enables the AI21 Fashions model to handle massive and complex information sets, thereby rendering it particularly valuable for tasks requiring comprehensive understanding and analysis of intricate data, such as lengthy documents or sophisticated data-intensive processes, while also significantly enhancing the retrieval stage of any RAG-based workflow. Unlike traditional Transformer-based language models that claim to support extensive context windows but often see their performance degrade with increasing input length, Jamba’s hybrid architecture ensures the mannequin’s quality remains uncompromised as context increases.
The Jamba 1.5 Mini’s 256K context window enables fashion designers to efficiently manage the equivalent of 800 pages of text content within a single glance, thereby streamlining their workflow and fostering creative collaboration. Several industries that can leverage Databricks include?
- Clients can leverage Jamba 1.5 Mini to swiftly condense lengthy reviews, contracts, and analysis papers into concise summaries. For financial institutions, the tools can condense earnings reports, dissect market trends from lengthy financial documents, or extract relevant information from regulatory submissions.
- By integrating diverse patient data streams, the mannequin enables healthcare providers to inform sophisticated medical decisions, generating personalized treatment recommendations.
- In Retailer-matched Answer Generation (RAG) methodologies, fashion can yield more accurate and contextually relevant responses to customer inquiries by considering a wider range of product information and buyer history.
Bitext offers a range of pre-trained, vertically specialized language models available on the Databricks Marketplace.
These fashion variants are derived from the Mistral-7B-Instruct-v0.2 mannequin, fine-tuned to develop chatbots, digital assistants, and copilots for retail banking applications, providing customers with prompt and accurate solutions for their financial inquiries. Here are the results:
The top-rated social buying and selling app struggled with alarmingly high drop-off rates during the customer onboarding process. The company revamped its onboarding process, transforming static tutorials into a conversational, intuitive, and personalized user experience that streamlined customer acquisition.
Bitext vertically integrated its artificial intelligence model and presented it to the client. Using the mannequin as a foundation, a knowledge scientist conducted initial fine-tuning by incorporating customer-specific data, including prevalent frequently asked questions. This step guaranteed that the mannequin comprehended the distinct requirements and vernacular specific to the target audience. This technology was successfully adopted through a collaboration between Databricks and Mosaic AI to deliver high-quality tuning results.
Once the Bitext mannequin had been thoroughly fine-tuned, it was successfully deployed using Databricks’ advanced AI model serving capabilities.
- The Unity-registered fine-tuned mannequin was cataloged.
- An endpoint was created.
- The mannequin was successfully deployed to the designated endpoint.
The groundbreaking partnership revolutionized customer interaction within the social finance industry, leading to a significant boost in buyer participation and loyalty. Due to the rapid acceleration enabled by the collaborative AI framework, our team successfully completed the entire project within a condensed two-week timeframe.
While generic fashion trends often require an abundance of training data, adopting a specialized model tailored to a specific industry significantly narrows down the information required to tailor it effectively. This enables rapid deployment of customized AI models. We’re excited about AI Model Sharing. Prospects have achieved up to 60% discounts on valuable resources, driven by fewer data scientists required and decreased computational needs, while also realizing up to 50% cost savings in operational disruptions through accelerated testing and deployment enabled by our specialized AI models available on the Databricks Marketplace.
— Antonio S. Valderrábanos , Founder & CEO, Bitext
| Excessive – In depth fine-tuning for sector & use case | What’s your prompt? | 60% | |
| Medium – Additional fine-tuning required | Low – Particular customization wanted | 30% | |
| 3-6 months | 1-2 months | 50-60% discount | |
| Scientists and computational resources are increasingly being overwhelmed by excessive data. | Low – Much less intensive | 40-50% | |
| Extended development timelines require longer integration and testing periods to ensure seamless system functionality. | Low – Quicker deployment | 50% |
Now that AI mannequin sharing has become widely accessible through General Availability (GA) for both Delta Sharing and the latest AI models available on the Databricks Marketplace, we invite you to:
