What's new and changed in watsonx.ai
Version 1.1.6
A new version of watsonx.ai was released in August 2024.
Version 1.1.6 is installed on IBM Cloud Pak® for Data 4.8.6 with the service operand version 8.6.0.
This release of the watsonx.ai service includes various fixes.
Version 1.1.5
A new version of watsonx.ai was released in April 2024.
Version 1.1.5 is installed on IBM Cloud Pak for Data 4.8.5 with the service operand version 8.5.0.
- New features
-
- Use Chat mode in Prompt Lab to inference models through conversational interactions
- Chat mode in Prompt Lab is a
simple chat interface that you can use to experiment with foundation models and simulate
question-answering or conversational interactions for chatbot and virtual assistant use cases. The
existing Structured and Freeform modes are useful when building few-shot or many-shot prompts for
tasks such as extraction, summarization, and classification. Chat mode remembers the context of the
conversation and each response builds on information from previous prompts.
For more information, see Prompt Lab.
- New watsonx.ai end-to-end use case tutorial and video
- Try the new tutorial to learn how to use watsonx.ai in an end-to-end use case starting from data preparation through to prompt engineering. For more information, see Try the watsonx.ai end-to-end use case
- Prompt-tune the granite-13b-instruct-v2 model
- You can tune the
granite-13b-instruct-v2
foundation model by using the Tuning Studio. For more information, see Tuning Studio. - Work with new foundation models in Prompt Lab
- You can now use the following foundation models for inferencing from the Prompt Lab in watsonx.ai:
jais-13b-chat
- The
jais-13b-chat
foundation model is provided by Inception, Mohamed bin Zayed University of Artificial Intelligence, and Cerebras Systems. This foundation model specializes in conversational tasks in Arabic and English. You can also use thejais-13b-chat
foundation model for general purpose tasks in the Arabic language, such as language translation between Arabic and English. llama2-13b-dpo-v7
- The
llama2-13b-dpo-v7
foundation model from Meta is trained to understand and generate Korean text. You can use thellama2-13b-dpo-v7
foundation model for general-purpose tasks in the Korean language, such as classification, extraction, question-answering, and for language translation between Korean and English.
For more information, see Supported foundation models.
Version 1.1.4
A new version of watsonx.ai was released in March 2024.
Version 1.1.4 is installed on IBM Cloud Pak for Data 4.8.4 with the service operand version 8.4.0.
- New features
-
- Bring your own foundation model to inference from watsonx.ai
- In addition to working with foundation models that are curated by IBM, you can now upload and deploy your own foundational models. After the models are deployed and registered with watsonx.ai, create prompts that inference the custom models from the Prompt Lab. To learn more about uploading custom foundation models, see Deploying custom foundation models. Also described is a list of supported GPU configurations, model architectures, and suggested matching hardware configurations.
- The watsonx.ai API is available for use
- You can now programmatically work with foundation models that are hosted in watsonx.ai by calling the watsonx.ai API directly. For more information, see the watsonx.ai API reference documentation.
- Prompt-tune the llama-2-13b-chat foundation model
- The Tuning Studio now supports
tuning the
llama-2-13b-chat
foundation model. Start by engineering prompts for the largerllama-2-70b-chat
model in the Prompt Lab to find effective prompt inputs for your use case. Then tune the smaller version of the Llama 2 model to generate comparable or better outputs with zero-shot prompts. For more information, see Tuning Studio. - A modification to the granite-13b-chat-v2 foundation model is available
- The latest version of the
granite-13b-chat-v2
is 2.1.0. The modification includes improvements that were gained by applying a novel AI alignment technique to the version 2.0.0 model. AI alignment involves using fine-tuning and reinforcement learning techniques to guide the model to return outputs that are as helpful, truthful, and transparent as possible. For more information, see the What is AI alignment? blog post from IBM Research.If you saved a prompt that inferences the granite-13b-chat-v2 foundation model as a template or notebook prior to this release, you might need to edit the asset to account for changes that were made to how the model handles white spaces.
- Work with new foundation models in Prompt Lab
- You can now use the following foundation models for inferencing from the Prompt Lab in watsonx.ai:
granite-8b-japanese
: A foundation model from the IBM Granite family that is trained to understand and generate Japanese text. You can use thegranite-8b-japanese
foundation model for general purpose tasks in the Japanese language, such as classification, extraction, question-answering, and for language translation between Japanese and English.granite-20b-multilingual
: A foundation model from the IBM Granite family that is trained for translation tasks in English, German, Spanish, French, and Portuguese.codellama-34b-instruct
: A programmatic code generation model from Code Llama that is based on Llama 2 from Meta. Use Code Llama to create prompts for generating code based on natural language inputs, and for completing and debugging code.mixtral-8x7b-instruct-v01-q
: A version of theMixtral 8x7B Instruct
foundation model from Mistral AI that is quanitzed by IBM. You can use this new model for general-purpose tasks, including classification, summarization, code generation, language translation, and more.
For more information, see Supported foundation models.
- Deprecated foundation models
- The following models are now deprecated and will be withdrawn in a future release:
granite-13b-chat-v1
granite-13b-instruction-v1
gpt-neox-20b
mpt-7b-instruct2
starcoder-15.5b
To avoid disruption when the withdrawal is completed, use a suitable alternative model as a replacement. For more information, see Foundation model lifecycle.
- Deprecated features
-
- Deprecated foundation models
- The following models are now deprecated and will be withdrawn in a future release:
granite-13b-chat-v1
granite-13b-instruction-v1
gpt-neox-20b
mpt-7b-instruct2
starcoder-15.5b
To avoid disruption when the withdrawal is completed, use a suitable alternative model as a replacement. For more information, see Foundation model lifecycle
Version 1.1.3
A new version of watsonx.ai was released in February 2024.
Version 1.1.3 is installed on IBM Cloud Pak for Data 4.8.3 with the service operand version 8.3.0.
- New features
-
- New Japanese-language Llama 2 foundation model for general-purpose tasks in Japanese
- The
elyza-japanese-llama-2-7b-instruct
foundation model provided by ELYZA, Inc. is available in watsonx.ai. Theelyza-japanese-llama-2-7b-instruct
model is a version of the Llama 2 model from Meta that was trained to understand and generate Japanese text.You can use this new model for general-purpose tasks. This model works well for Japanese-language classification and extraction and for text translation between Japanese and English.
For more information, see Supported foundation models.
- Deploy prompt templates for foundation models
- You can now save a prompt template as a reusable asset and use it to generate output from a foundation model. Deploy the prompt template to a deployment space to test it or to retrieve the endpoint to use in an application. For more information, see Deploying a prompt template.
Version 1.1.1
A new version of watsonx.ai was released in December 2023.
Version 1.1.1 is installed on IBM Cloud Pak for Data 4.8.1 with the service operand version 8.1.0.
- New features
-
- Explore the risks of working with AI
- You can now explore some of the risks of working with generative AI, foundation models, and machine learning models. Read about risks for privacy, fairness, explainability, value alignment, and other areas. For more information, see AI risk atlas.
- Use variables to build reusable prompts
- Add flexibility to your prompts with prompt variables. Prompt variables function as placeholders in the static text of your prompt input. You can replace this placeholder with text dynamically at inference time. You can save prompt variable names and default values in a prompt template asset to reuse yourself or share with collaborators in your project. For more information, see Building reusable prompts.
- Apply AI guardrails
- You can now enable AI guardrails in the Prompt Lab. AI guardrails remove potentially harmful text from the input to and output from foundation models. Harmful text can include hate speech, abusive language, and profanity. For more information, see Prompt Lab.
- Introducing the Tuning Studio
- Use the Tuning Studio to tailor the output that a foundation model
returns to better meet your needs. When you tune a model, you influence the content that is included
in the model’s output and how the content is formatted. For example, if you plan to add the step of
inferencing a foundation model into a business workflow, then you need to:
- Know what to expect from the foundation model.
- Be able to retrieve the output in a reliable format.
With the Tuning Studio, you can apply the prompt tuning method to the Google
flan-t5-xl-3b
model. For more information, see Tuning Studio. - Work with new foundation models in Prompt Lab
- You can now use the following foundation models for inferencing from the Prompt Lab in watsonx.ai:
flan-t5-xl-3b
granite-13b-chat-v2
granite-13b-instruct-v2
llama-2-13b-chat
- Add sample prompts to notebooks
- You can now add various sample prompts for specific models into your notebooks. To add a sample prompt, click the Code snippets icon (), select Prompt Engineering, and browse the various categories to find a sample prompt. When you select a prompt, click Insert code to cell to insert the prompt into your notebook.
Version 1.1.0
A new version of watsonx.ai was released in November 2023.
Version 1.1.0 is installed on IBM Cloud Pak for Data 4.8.0 with the service operand version 8.0.0.
This release includes the following changes:
- New features
-
Use the watsonx.ai service to transform your business with generative AI. With watsonx.ai, you can try, validate, and deploy a range of foundation models, including:
- The IBM Granite series of foundation models. The Granite family includes decoder-only models that are instruction-tuned to be good at enterprise tasks, such as summarization, extraction, classification, and more.
- Third-party models including Llama 2 Chat for dialogue and StarCoder for code generation.
- Other popular open source foundation models that are available from Hugging Face.
You can interact with foundation models in two ways:- Programmatically by using the Python library
- Directly by using the Prompt Lab, an intuitive tool for engineering prompts