Follow this comprehensive guide to successfully deploy IBM Granite Guardian Models for your enterprise.
IBM Granite
IBM family of open, dependable, and high-performing AI models, IBM Granite, is designed for corporate use and scales your AI applications.
IBM Granite Models
Granite 3.1 Language models
Base and instruction-tuned language models for RAG, content creation, text analytics and extraction, text summarisation, agentic workflows, and classification.
Granite-3.1-8B-Instruct
Granite-3.1-8B-Instruct is an 8B parameter long-context instruct model that was refined from Granite-3.1-8B-Base by combining internally gathered synthetic datasets designed for long context issue resolution with open source instruction datasets under permissive licenses.
Granite-3.1-2B-Instruct
Granite-3.1-2B-Instruct is a 2B parameter long-context instruct model that was refined from Granite-3.1-2B-Base by combining internally gathered synthetic datasets designed for long context issue solving with open source instruction datasets under permissive licenses.
Granite-3.1-3B-A800M-Instruct
Granite-3.1-3B-A800M-Instruct is a 3B parameter long-context instruct model that was refined from Granite-3.1-3B-A800M-Base utilising both internally gathered synthetic datasets specifically designed for long context issue solving and open source instruction datasets with permissive licenses.
Granite-3.1-1B-A400M-Instruct
Granite-3.1-1B-A400M-Instruct is an 8B parameter long-context instruct model that was refined from Granite-3.1-1B-A400M-Base using a combination of internally gathered synthetic datasets designed for long context issue solving and open source instruction datasets with permissive licenses.
For all Granite 3.1 Language models, a variety of methods, including as supervised fine-tuning, model alignment through reinforcement learning, and model merging, are used to create this model in an organised conversation format.
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese are among the languages that are supported. Beyond these 12 languages, users can adjust Granite 3.1 models.
Intended Use: The model can be used to create AI assistants for a variety of fields, including business applications, and is made to react to generic instructions.
Capabilities
- Synopsis
- Classification of texts
- Extraction of text
- Answering questions
- Tasks involving Retrieval Augmented Generation (RAG) code
- Tasks that call functions
- Use cases for multilingual dialogue
- Long-context jobs, such as long document QA and long document/meeting summarisation, etc.
Granite for the code
Code produced in 116 different programming languages was used to train decoder-only models for code generating tasks, such as code generation, code explanation, and code editing.
Granite for time series
It is lightweight, pre-trained for time-series forecasting, and designed to function well with a variety of hardware setups.
Granite Guardian
Protect AI with Granite Guardian, which performs very well in more than 15 safety standards while guaranteeing enterprise data protection and reducing risks across a range of user prompts and LLM replies.
Granite Guardian 3.1 8B
A refined version of Granite 3.1 8B Instruct, Granite Guardian 3.1 8B is intended to identify potential hazards in prompts and responses. Along a number of important dimensions included in the IBM AI Risk Atlas, it can aid in risk detection. It is trained on distinct data that includes both synthetic data gleaned from internal red-teaming and human annotations. On common benchmarks, it performs better than other open-source models in the same field.
Use
Granite Guardian is helpful for use cases involving risk detection that may be used to a variety of business applications.
identifying potential hazards of harm in the model response or prompt language (as guardrails). Since the former analyses user-supplied language and the latter evaluates model-generated text, these offer two essentially distinct use cases.
The guardian model evaluates three primary factors in the RAG (retrieval-augmented generation) use-case: answer relevance (whether the response directly answers the user’s query), groundedness (whether the response is accurate and faithful to the provided context), and context relevance (whether the retrieved context is relevant to the query).
Granite Guardian assesses intermediary steps for syntactic and semantic hallucinations in order to detect function calling danger in agentic workflows. This involves determining whether function calls are legitimate and identifying false information, especially when translating queries.
Risk Definitions
The model is specifically made to identify different kinds of dangers in messages sent by users and assistants. This contains the following particular dangers as well as an umbrella category called “Harma” that is intended to detect information that is widely acknowledged to be hazardous in unconventional ways.
- Content deemed generally damaging is referred to as harm.
- Prejudice based on identification or traits is known as social bias.
- Jailbreaking is the intentional use of AI manipulation to produce offensive, undesirable, or dangerous content.
- Violence: information that encourages sexual, mental, or bodily harm.
- Profanity is the use of insulting or offensive language.
- Sexual Content: sexually suggestive or explicit content.
- Behaviour that transgresses moral or legal norms is considered unethical.
An innovative use of the concept is the evaluation of hallucinatory hazards in a RAG pipeline. These consist of
- Relevance of Context: The context that was retrieved is irrelevant to resolving the user’s demands or providing an answer to their query.
- Groundedness: assertions or facts in the assistant’s response that are not backed up by or in conflict with the context that was given.
- Answer Relevance: The user’s input is not adequately addressed or responded to by the assistant.
Additionally, the model may identify dangers in agentic workflows, including
- Function Calling Hallucination: depending on the user’s question and the tool at hand, the assistant’s response may comprise function calls with syntactic or semantic problems.
Granite Guardian 3.1 2B
A refined Granite 3.1 2B Instruct model, Granite Guardian 3.1 2B is intended to identify hazards in prompts and responses. Along a number of important dimensions included in the IBM AI Risk Atlas, it can aid in risk detection. It is trained on distinct data that includes both synthetic data gleaned from internal red-teaming and human annotations. On common benchmarks, it performs better than other open-source models in the same field.
Risk Definition
Same as Granite Guardian 3.1 8B
Granite for geospatial data
Together, NASA and IBM used extensive satellite and remote sensing data to develop an AI Foundation Model for Earth Observations.
Granite embedding models
intended to greatly improve comprehension of user intent and raise the relevance of sources and information in answer to a query.
Comparisons
Specialised use cases were given priority in earlier Granite model iterations. Along with providing even more effectiveness in those areas, IBM Granite 3.0 models meet and often surpass the overall performance of top open weight LLMs in terms of both academic and business standards.
Why Granite?
Open
Select the appropriate model, open-sourced under Apache 2.0, with parameters ranging from sub-billion to 34 billion.
Performer
Performance should never be compromised for price. Granite performs better than similar models1 in a range of business tasks.
Reliable
Create responsible AI with a full suite of IP protection, transparency, and risk and harm detection features.
Build with Granite
Use Red Hat Enterprise Linux AI with Watsonx to put open-source Granite models into production, giving you the resources and assistance you need to implement AI at scale with assurance. With features like tool-calling, 12 languages, multi-modal adapters (coming soon), and more, you can build more quickly.