Clarifai 10.8: Supercharge AI Models with Advanced Concept Mapping | Clarifai Blog

This blog post focuses on new features and improvements. For a comprehensive list, including bug fixes, please see the release notes.

Advanced Concept Management

A concept represents an entity, similar to a "tag" or "keyword," and is also referred to as a "class" in machine learning. Concepts can be used to annotate inputs that contain the associated entity or added to a model for entity recognition. The data tied to these concepts helps the model learn and understand the entity.

On the Clarifai Platform, you can upload your inputs, create concepts, and manage various tasks within your app. We’ve now introduced a new way to define concept relations in Clarifai, allowing you to define hierarchical and semantic links between concepts. This helps build robust knowledge graphs and enabling more advanced computer vision and natural language processing capabilities.

By mapping concepts as synonyms, hyponyms, or hypernyms, you can create context-aware models that deliver more accurate results. This is especially useful for managing data, building search engines, and organizing large datasets, ultimately improving the effectiveness of AI applications.

Hyponym (Subtype Relation): Represents a "kind of" relationship, like "Dog" being a kind of "Animal". This hierarchy helps models generalize or specialize in identifying concepts. Usage: Hyponyms refine search results; a search for "Animal" might return "Dog", "Cat", and other animals.

Hypernym (Supertype Relation): The opposite of a hyponym, it indicates a broader category or parent concept such as "Animal" being a hypernym for "Dog" or "Animal" is a parent of "Dog". Usage: Hypernyms group specific entities under broader categories, helping in data organization.

Synonym (Equivalent Relation): Links concepts with the same or similar meanings, like "Puppy" and "Pup". Usage: Synonyms ensure that different terms for the same concept are treated as equivalent, improving search accuracy.

Below is an example of how to add and manage concept relations in the Clarifai app using the Python SDK.

Explore this guide and included notebook to learn how to add concept relations to your data, and follow along step by step.

Published new model: Prompt-Guard-86M

LLM-powered applications can be vulnerable to prompt attacks, where malicious prompts are designed to manipulate the model's behavior against the developer's intentions.

There are two different prompt attacks:

• Prompt Injections: These are inputs that take advantage of combining untrusted data from third parties or users into a model's context, causing the model to follow unintended instructions.

• Jailbreaks: These are malicious instructions intended to override a model's built-in safety and security features.

Prompt-Guard-86M is a multilingual classifier model designed to detect and prevent these prompt injections and jailbreak attacks in LLM-powered applications. The model is trained on a comprehensive corpus of attack data to detect both explicitly malicious prompts and those containing subtle injected inputs.

As a multi-label model, it categorizes inputs into three distinct classes: benign, injection, and jailbreak, thereby providing a robust mechanism to mitigate risks in LLM-powered applications.

The model is now available on the Clarifai Platform, We have some pre built examples to get you started. Try out Prompt-Guard-86M!

Input-Viewer

Added option to download original asset from Input-Viewer

• You can now download the original asset directly from the Input-Viewer page. If you right-click on the canvas area, a button will appear, which allows you to download the original input. Previously, only masked versions could be downloaded.

Enhanced annotation identification in Input-Viewer

• When hovering over or selecting an annotated object on the canvas of the Input-Viewer, the corresponding annotation in the right sidebar is now highlighted. This makes it easier for users to quickly identify the annotation for deletion, editing, or other purposes.

Deprecated some agent system operators

Agent system operators are fixed-function operators that act as "non-trainable models". They help you connect, direct, and network your models in a workflow.

We have deprecated the following agent system operators: Custom Code Operator, AWS Lambda, Isolation Operator, Tesseract Operator, Neural Tracker and Neural Lite Tracker.

Read more about Agent System Operators and how you can use them with workflows for various use cases here.

Additional changes

• Enabled Secure Data Hosting(SDH) for some model features: We now serve the model version train logs and model version export capabilities via secure-data-hosting instead of pre-signed urls.

• Added missing buttons in reviewer field for labeling tasks: We’ve improved the reviewer selection field by adding two new buttons: “Add/Remove all collaborators” and “New collaborator.” These options streamline the process of managing reviewers when creating a new labeling task.

• Introduced automatic resizing of bounding boxes/polygons: When annotating and drawing a bounding box or polygon, if the shape is stretched beyond the main canvas area on any side, it will now automatically resize and adjust to fit within the canvas edge on the same side.

• Removed scrollbars on smaller screens: We’ve removed horizontal and vertical scrollbars from the left menu on smaller screen sizes. This adjustment ensures a cleaner, more user-friendly interface when the window height is small, which improves the overall user experience.

Ready to start building?

On the Community Platform, you can upload your inputs, create concepts, and manage various tasks within your app. Check out our quick-start guide on the Community Platform. Sign up here to get started!

If you have any questions, send us a message in our Community Discord channel. Thanks for reading!