Amazon has announced the launch of its new fully managed service, Amazon Bedrock, which provides a selection of high-performance base models from leading AI companies such as AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon itself. This service, accessible through a single API, focuses on enabling the building of generative artificial intelligence applications with high standards of security, privacy, and responsibility.
Base models are especially useful in common generative AI use cases, such as chatbots, virtual assistants, and conversational search, among others. To optimize the response of these models, a technique known as Retrieval-Augmented Generation (RAG) provides specific contexts, and it is recommended to finely tune the models to align them with the industry or brand voice and vocabulary.
A standout feature of Amazon Bedrock is the support for importing custom models, allowing users to incorporate previously fine-tuned models from other environments such as Amazon SageMaker, Amazon EC2, and on-premises locations, directly into Amazon Bedrock. This functionality, currently in preview phase, facilitates the integration of custom and finely tuned models, improving the adaptability and performance of generative AI applications.
Key features of the Custom Model Import service include:
1. Ability to import fine-tuned models to leverage Amazon Bedrock’s fully managed serverless capabilities.
2. Support for model architectures like Llama 2, Llama 3, Flan, and Mistral, with FP32, FP16, and BF16 precisions and future quantizations.
3. Capability to run the import process with model weights stored on Amazon S3.
4. Direct integration with Amazon SageMaker models using Amazon Resource Names (ARN) references.
5. Automatic scalability of the model based on traffic patterns, optimizing costs by reducing resources when not in use.
The process of importing and evaluating a finely tuned model involves detailed steps such as tuning the model using SageMaker, importing it into Amazon Bedrock, and evaluating it using the SageMaker FMEval library. For example, tuning a Mistral model using the OpenOrca dataset and its subsequent evaluation are described, ensuring the accuracy and efficiency of the model based on specific user needs.
To illustrate the solution, the article includes workflow diagrams and code examples for each of the steps, from preparing the dataset to evaluating the imported model. In this way, Amazon Bedrock not only facilitates the import of models but also their scalable and secure implementation.
With Amazon Bedrock, Amazon expands its AI offering, providing developers with advanced tools to efficiently build and scale generative AI applications.
via: MiMub in Spanish