AWS is pleased to introduce the Amazon Titan Text Premier, a new model in the Amazon Titan line that is currently offered in Amazon Bedrock.
Titan Text Premier is the newest large language model (LLM) in the Amazon Titan family of models, which expands your model selection even more within Amazon Bedrock, and comes after Amazon Titan Text Lite and Titan Text Express. In Bedrock, you can now select among the following Titan Text models:
Titan Text Premier
The most sophisticated Titan LLM for text-based enterprise applications is Titan Text Premier. It has been especially optimised for enterprise use cases, such as constructing agent-based applications with Knowledge Bases and Agents for Amazon Bedrock and Retrieval Augmented Generation (RAG), with a maximum context length of 32K tokens. Titan Text Premier is most suited for English-language jobs, however it has pre-trained on multilingual text data like all Titan LLMs do. With your own data in Amazon Bedrock, you may further fine-tune (preview) Titan Text Premier to create applications that are unique to your domain, company, brand, and use case. I’ll go into more detail on the model’s performance and highlights in the parts that follow.
Titan Text Messenger
For a variety of activities, including conversational discussion and open-ended text production, Titan Text Express is perfect. The maximum context length for the model is 8K tokens.
Titan TextLite
Titan Text Lite may be fine-tuned for activities like article summarising and copywriting because it is very customisable, fast, and optimised for speed. The maximum context length for the model is 4K tokens.
Let’s talk more in-depth about Titan Text Premier now.
Highlights of the Amazon Titan Text Premier model
Titan Text Premier has been fine-tuned to incorporate appropriate artificial intelligence (AI) techniques and optimised for high-quality RAG and agent-based applications.
Enhanced for RAG and agent-oriented software
In response to client feedback indicating that RAG is a crucial component in developing generative AI applications, Titan Text Premier has been specifically optimised for RAG and agent-based applications. The model training data has been optimised for interaction with Knowledge Bases and Agents for Amazon Bedrock, and includes examples for activities like conversational chat, Q&A, and summarization. As part of the optimisation, the model is trained to manage the subtleties of these properties, like their unique prompt formats.
Integrating Amazon Bedrock Knowledge Bases for high-quality RAG
Superior RAG by use of Knowledge Base integration for Amazon Bedrock You may safely link your company’s data for RAG to the foundation models (FMs) in Amazon Bedrock by using a knowledge base. Titan Text Premier with Knowledge Bases is now an option for implementing tasks like question-answering and summarising over proprietary data that belongs to your firm.
Task automation with integration with Agents for Amazon Bedrock
Using Titan Text Premier with Agents for Amazon Bedrock, you can also build bespoke agents that can carry out multistep processes across various enterprise systems and data sources. You can use agents to automate processes for your clients, both internal and external, such handling insurance claims or retail order management.
Amazon Titan Text Premier customisation (preview)
Amazon Titan Text Premier’s custom fine-tuning (preview) By giving Titan Text Premier your own task-specific labelled training dataset in Amazon Bedrock, you may fine-tune the model and improve accuracy. Titan Text Premier customisation allows you to further specialise your model and produce distinctive user experiences that are consistent with your business’s voice, style, brand, and offerings.
Built responsibly
Constructed with responsibility, Amazon Titan Text Premier integrates reliable, safe, and secure procedures. The model’s performance is documented in the AWS AI Service Card for Amazon Titan Text Premier, covering important responsible AI benchmarks such as robustness, safety, and fairness. Additionally, the model interfaces with Guardrails for Amazon Bedrock, allowing you to add more security according to your application’s needs and ethical AI guidelines. Customers who use Amazon Titan models responsibly are protected by Amazon against accusations that their outputs or the models themselves violate third-party copyrights.
Model performance of the Amazon Titan Text Premier
Titan Text Premier is designed to provide organisations with a wide range of useful information. The evaluation results for public benchmarks, which measure important abilities including instruction following, reading comprehension, and multistep reasoning versus models with comparable prices, are displayed in the following table. Titan Text Premier’s robust performance on a variety of demanding benchmarks demonstrates its excellent price-performance and ability to handle a broad range of use cases in enterprise applications. Higher scores are better for all benchmarks stated below.
Capability | Benchmark | Description | Amazon | OpenAI | |
---|---|---|---|---|---|
Titan Text Premier | Gemini Pro 1.0 | GPT-3.5 | |||
General | MMLU (Paper) | Representation of questions in 57 subjects | 70.4% (5-shot) | 71.8% (5-shot) | 70.0% (5-shot) |
Instruction following | IFEval (Paper) | Instruction-following evaluation for large language models | 64.6% (0-shot) | not published | not published |
Reading comprehension | RACE-H (Paper) | Large-scale reading comprehension | 89.7% (5-shot) | not published | not published |
Reasoning | HellaSwag (Paper) | Common-sense reasoning | 92.6% (10-shot) | 84.7% (10-shot) | 85.5% (10-shot) |
DROP, F1 score (Paper) | Reasoning over text | 77.9 (3-shot) | 74.1 (Variable Shots) | 64.1 (3-shot) | |
BIG-Bench Hard (Paper) | Challenging tasks requiring multistep reasoning | 73.7% (3-shot CoT) | 75.0% (3-shot CoT) | not published | |
ARC-Challenge (Paper) | Common-sense reasoning | 85.8% (5-shot) | not published | 85.2% (25-shot) |
It should be noted that benchmarks use a combination of few-shot and zero-shot prompting to assess model performance. A few-shot prompting technique involves giving the model several real-world examples (three for 3-shot, five for 5-shot, etc.) of how to complete a certain task. This illustrates the model’s capacity for in-context learning, or learning from examples. On the other hand, zero-shot prompting lets you assess a model’s performance without giving it any examples, depending just on its prior knowledge and comprehension of generic language.
Use Amazon Titan Text Premier to get started
Go to the Amazon Bedrock panel and select Model access from the bottom left pane to permit access to Amazon Titan Text Premier. To enable access to Amazon Titan Text Premier, select the Manage model access button located in the upper right corner of the Model access overview page.
In the Bedrock console, select Text or Chat from the Playgrounds menu on the left side pane to utilise Amazon Titan Text Premier. Next, select Titan Text Premier as the model and Amazon as the category by clicking on Select model. You can load examples to explore the model. One of those instances that highlights the model’s chain of thought (CoT) and reasoning ability is displayed in the screenshot that follows.
You can obtain a code example demonstrating how to use the AWS Command Line Interface (AWS CLI) to execute the model with the current example prompt by selecting View API request. Using the AWS SDKs, you can also access Amazon Bedrock and the available models. You will utilise the AWS SDK for Python (Boto3) in the ensuing example.
Currently accessible
As of right now, the AWS US East (North Virginia) Region offers Amazon Titan Text Premier. Amazon Titan Text Premier custom fine-tuning is now available in preview in the AWS US East (North Virginia) Region. See the complete list of regions for upcoming changes. Go to the Amazon Titan product website to find out more about the Amazon Titan model line. See the Amazon Bedrock pricing page for specific pricing information.