This weblog put up focuses on new options and enhancements. For a complete checklist, together with bug fixes, please see the launch notes.
Management Middle
We’re excited to introduce the brand new Clarifai Management Middle, the unified dashboard, a single pane of glass to watch every part taking place inside your account on the platform. It serves as the only supply of fact for varied data dimensions, enabling you to make knowledgeable selections primarily based on knowledge from a number of sources.
Management Middle helps streamline the administration of your Clarifai operations by consolidating all actions right into a single interface, minimizing the necessity to change between completely different instruments or home windows. It lets you observe your utilization of Clarifai’s sources utilizing charts, graphs, tables, and different instruments, and centralizes monetary knowledge and useful resource utilization into one interface.
Moreover, it affords the flexibility to rework uncooked knowledge reminiscent of mannequin predictions, search operations, saved inputs, and mannequin coaching hours into actionable insights. You’ll be able to observe how your groups operate and the way it impacts operations and spend, and likewise audit any user-initiated occasion to hint its origin.
Take a look at the fast tutorial under to be taught extra about Management Middle. You may also discover the detailed data on our docs right here.
Revealed new fashions
- OpenAI o1-preview: A reasoning-focused AI mannequin designed for complicated problem-solving in math, coding, and science. The mannequin scored within the 89th percentile on Codeforces programming competitions and achieved 83% on Worldwide Arithmetic Olympiad qualifying questions, in comparison with GPT-4o’s 14%. It’s able to fixing PhD-level issues in physics, biology, and chemistry.
- OpenAI o1-mini: A smaller mannequin that’s 80% cheaper than o1-preview and delivers nearly the identical efficiency for STEM challenges.
- Llama-3.2-11B-Imaginative and prescient-Instruct: A 11B parameter open supply multimodal LLM by Meta designed for visible reasoning, picture captioning, and VQA duties, supporting textual content and picture inputs.
- Llama-3.2-3B-Instruct: A multilingual, instruction-tuned LLM optimized for dialogue and textual content era duties.
Introducing enhanced mannequin add capabilities [Private Preview]
We launched a brand new technique for mannequin uploads to the platform. This integration consists of:
- A brand new and extra environment friendly mannequin add course of using Clarifai’s runners.
- Migration of related information from runners-python to applicable places inside the clarifai-python repository.
Study extra about it right here. This function is presently in Personal Preview. We might love so that you can strive it out and supply your suggestions. To take action be happy to enroll in the Personal Preview right here.
Enhanced mannequin add capabilities with HuggingFace integration
We have expanded our mannequin add performance to enhance integration with HuggingFace and improve idea dealing with:
- Launched a HuggingFaceLoader utility for seamless checkpoint downloads and idea fetching from HuggingFace for classification and detection fashions.
- Applied new strategies in model_upload for environment friendly checkpoint downloads and idea dealing with.
- Up to date the mannequin add course of to make use of the brand new HuggingFaceLoader for HuggingFace-related fashions.
- Modified the command-line interface to make use of –model_path as an alternative of -folder for improved readability.
Extra modifications
- We’ve launched “Collaborations” as a top-level hyperlink, just like “My Apps” and “Group,” making it simpler for customers to entry apps they’ve been invited to collaborate on.
- Beforehand, modifying a RAG Prompter mannequin in a public workflow would unexpectedly create a brand new non-public mannequin and model, inflicting save failures as a result of incompatibility of personal fashions in public workflows. This additionally led to workflow execution errors. We have resolved this by guaranteeing that modifications to current nodes reuse the unique mannequin, both creating a brand new model or updating the present one as applicable. This repair maintains workflow integrity, prevents unintended privatization, and eliminates associated execution errors.
- Beforehand, Org Contributors have been in a position to view the pending members tab. Since inviting members is an Admin-only operate, we now limit Org Contributors, Org Customers, and Staff Contributors from accessing the pending invites checklist.
Prepared to start out constructing?
Discover our Quickstart tutorial to study Management Middle. Enroll right here to get began!
We have additionally printed a public roadmap to present you a view of our upcoming plans and to assemble your priceless suggestions. You’ll be able to discover the options we’re contemplating, vote for those you need most, or submit new concepts. Take a look at the roadmap right here.
When you’ve got any questions, ship us a message on our Group Discord channel. Thanks for studying!