API
Added inference parameters customization and versatile API key choice
- Now you can configure inference parameters equivalent to temperature, max tokens, and extra, relying on the precise mannequin you might be utilizing, for each text-to-text and text-to-image generative duties. This empowers you to customise and fine-tune your mannequin interactions to higher fit your particular person wants.
Bug Fixes
- Fastened a difficulty the place the ListWorkflows endpoint fetched workflows from deleted apps. Beforehand, workflows from deleted functions have been fetched within the response of the ListWorkflows endpoint. We mounted the problem.
Python SDK
Added inference parameters customization and versatile API key choice
- The SDK helps the next textual content information codecs: plain textual content, CSV, and TSV (in roadmap).
- It helps the next picture information codecs: JPEG, PNG, TIFF, BMP, WEBP, and GIF.
- You could have the pliability to add information from both URLs or native storage, whether or not it is inside a single folder or unfold throughout a number of sub-folders.
- You’ll be able to add textual content information enriched with metadata and textual content classification annotations.
- You too can add picture information enriched with annotations, geoinformation, and metadata info.
- You’ll be able to effectively create apps designed for textual content mode and picture mode use instances.
- The SDK ensures clear progress monitoring throughout every add and promptly studies any encountered errors.
Built-in YAML workflow interfaces with proto endpoints
- Within the SDK, the workflow interfaces are actually represented in YAML format, however they’re seamlessly built-in with proto endpoints for environment friendly communication and interplay.
Enhanced the SDK’s foundational construction
- The SDK now helps authentication solely utilizing Private Entry Tokens (PAT) and UserIDs.
- We plan to supply helper utilities for wealthy textual content formatting and printing to facilitate a user-friendly interplay.
- We’re working to always add a strong error-handling system that covers varied eventualities.
New Printed Fashions
Printed a number of new, ground-breaking fashions
- Wrapped Falcon-40B-Instruct, a causal decoder-only Giant Language Mannequin (LLM) constructed by TII primarily based on Falcon-40B and fine-tuned on a combination of Baize.
- Wrapped WizardLM-70B, a mannequin fine-tuned on the Llama2-70B mannequin utilizing Evol+ strategies; it delivers excellent efficiency.
- Wrapped CodeLlama-13B-Instruct-GPTQ, CodeLlama-34B-Instruct-GPTQ, and CodeLlama-7B-GPTQ. Code Llama is a household of code-focused LLMs, constructed upon Llama 2. These fashions excel at filling in code, dealing with in depth enter contexts, and following programming directions with out prior coaching for varied programming duties.
- Wrapped WizardLM-13B, an LLM fine-tuned on the Llama-2-13b mannequin utilizing the Evol+ strategy; it delivers excellent efficiency.
- Wrapped WizardCoder-15B, a code-based LLM that has been fine-tuned on Llama 2 utilizing the Evol-Instruct methodology and has demonstrated superior efficiency in comparison with different open-source and closed-source LLMs on distinguished code technology benchmarks.
- Wrapped WizardCoder-Python-34B, a code-based LLM that has been fine-tuned on Llama 2. It excels in Python code technology duties and has demonstrated superior efficiency in comparison with different open-source and closed-source LLMs on distinguished code technology benchmarks.
- Wrapped Phi-1, a high-performing 1.3 billion-parameter text-to-code language mannequin, excelling in Python code technology duties whereas prioritizing high-quality coaching information.
- Wrapped Phi-1.5, a 1.3 billion parameter LLM that excels at complicated reasoning duties and was skilled on a high-quality artificial dataset.
- Wrapped OpenAI’s GPT-3.5-Turbo-Instruct, an LLM designed to excel in understanding and executing particular directions effectively. It excels at finishing duties and offering direct solutions to questions.
- Wrapped Mistral-7B-Instruct, a state-of-the-art 7.3 billion parameter language mannequin, outperforming Llama2-13B in a number of NLP benchmarks, together with code-related challenges.
LlamaIndex
We now help integrating with the LlamaIndex information framework for varied use instances, together with:
- Ingesting, structuring, and accessing exterior information for the LLMs you fine-tune on our platform. This lets you enhance the accuracy of your LLM functions with personal or domain-specific information.
- Storing and indexing your information for varied functions, together with integrating with our downstream vector search and database companies.
- Constructing an embeddings question interface that accepts any enter immediate and leverages your information to supply knowledge-augmented responses.
Fashions
Added skill to fine-tune text-to-text fashions
- Superior mannequin builders can now additional customise the habits and output of the text-to-text fashions for particular textual content technology duties. They will prepare the fashions on particular datasets to adapt their habits for explicit duties or domains.
[Enterprise Preview] Fastened a difficulty with the Leaderboard, which organizes fashions primarily based on their analysis outcomes
- Beforehand, for visible detection fashions, the displayed metrics included classification metrics like ROC, as an alternative of the anticipated metric, which is Imply Common Precision (MAP). The Leaderboard now accurately shows MAP, identical to within the mannequin viewer’s model desk.
Bug Fixes
- Launched multi-label textual content classification with Hugging Face textual content classification pipeline. Beforehand, there was a difficulty with multi-label textual content classification utilizing the Hugging Face textual content classification pipeline. Whereas it labored effectively for multi-class classification, it didn’t carry out accurately for multi-label classification. We mounted the problem.
- Fastened points with the mannequin model particulars web page.
- Fastened a difficulty that prevented the mannequin model particulars web page from displaying any output fields.
- Fastened a difficulty that induced the mannequin model particulars desk web page to crash unexpectedly.
- Fastened a difficulty the place the mannequin model particulars web page displayed incorrect parameter values when switching between completely different mannequin variations.
- Fastened a difficulty the place the mannequin model particulars web page displayed fields irrelevant to the precise mannequin kind. Pointless fields unrelated to the chosen mannequin kind are actually not seen.
- Fastened a difficulty the place the Mannequin Viewer web page incorrectly rotated photos in particular group apps. Beforehand, when photos with a portrait orientation (peak higher than width) have been submitted to some apps, the expected detection bins appeared misaligned or incorrect. The Mannequin Viewer web page now accurately rotates photos with portrait orientation, aligning detection bins precisely for improved consumer expertise and detection accuracy.
- Fastened a difficulty the place the mannequin ID disappeared after deciding on the Base embed_model within the mannequin creation course of. Beforehand, customers encountered a difficulty the place their beforehand entered mannequin ID would disappear upon deciding on a base embed mannequin when creating a brand new mannequin. We mounted the problem.
Modules
Eliminated collaborative functions from the apps drop-down record throughout the Set up Module modal when an org consumer is chosen
- Since group customers should not have collaborative apps or have interaction in collaborations, we have now optimized the interface by eradicating collaborated apps within the Set up Module modal particularly for this consumer group. This additionally mounted the app failure bug.
Enhanced the default app choice within the Set up Module modal
- The Set up Module modal now routinely selects your present app because the default vacation spot to put in your chosen module.
Improved the search performance for modules
- This refinement within the search function affords a extra streamlined and productive expertise when in search of particular modules or associated content material.
Launched a useful addition to the module’s viewer web page—an easy-to-use copy icon
- Now you can click on the copy icon positioned within the Newest Model ID part to swiftly receive the whole newest model URL for a module. This URL means that you can expedite the module set up course of on your utility.
Enter-Supervisor
Added skill to view and tag textual content inputs throughout the Enter-Viewer
- We’ve designed a complete UI that means that you can interactively create and handle annotations for textual content inputs.
Bug Fixes
- Improved keyboard shortcut performance in Enter-Viewer. Within the Enter-Viewer, you should utilize keyboard shortcuts (hotkeys), equivalent to H, V, P, and B, to change between annotation instruments.
- Beforehand, when opening the mannequin selector or some other factor with a popover whereas within the Enter-Viewer, typing any of the hotkeys would unintentionally change instruments within the background. We mounted the problem, and the hotkeys don’t now set off occasions that have an effect on instruments within the background.
- Enabled error-free transition to annotation mode in Enter-Viewer. Beforehand, there was an error instantly after switching to annotation mode within the Enter-Viewer. We mounted the problem.
- Fastened difficulty stopping collaborators from creating annotations in Enter-Viewer. Collaborators can now actively take part within the annotation course of, contributing to a extra collaborative and environment friendly workflow.
- Improved the Enter-Viewer’s URL dealing with to make sure a extra seamless navigation expertise. Beforehand, there was a difficulty the place the enter ID a part of the route parameters was not utilized successfully for rendering, resulting in undesired redirects. We mounted the problem.
Invites
- Now you can seamlessly evaluation and settle for pending Group invites instantly throughout the Portal, eliminating the necessity to rely solely on electronic mail for this function. That is beneficial for each our SaaS providing and our on-premise deployments, because it caters to eventualities the place electronic mail help could also be restricted or unavailable.
Group Settings and Administration
Bug Fixes
- Fastened a difficulty with displaying incorrect menu objects within the cellular view. Beforehand, whenever you switched out of your private consumer account to a company account by clicking your profile icon on the top-right nook of the navigation menu bar, the drop-down record displayed your private particulars as an alternative of the group’s info. We mounted the problem.
- Fastened a difficulty the place a company contributor couldn’t entry public workflows. Group contributors (group members) can now fetch a whole record of all public workflows with none hindrances.
- Enabled group contributors to view out there vacation spot apps in Set up Module Modal. We mounted a difficulty that beforehand hindered group contributors from viewing the record of obtainable vacation spot apps throughout the Set up Module modal when trying to put in a module.
- Fastened a difficulty with looking for group members. Beforehand, whenever you used a number of parameters, equivalent to each first identify and final identify collectively, to seek for members whose first identify or final identify matched the question textual content, it might end in an error throughout the response. We mounted the problem.
- Fastened a difficulty the place collaborators with the group contributor position did not record mannequin evaluations. Beforehand, collaborators attempting to record mannequin evaluations encountered an error message stating “Inadequate scopes”, regardless of the expectation that the position would grant them entry to view evaluations. We mounted the problem.
Apps
Enhanced App Settings web page to show particulars of collaborators
- Now, you possibly can simply entry and consider the small print of your collaborators, together with their ID, first and final identify, electronic mail, job position, and the date they have been added to the app.
Bug Fixes
- Fastened a difficulty the place the app thumbnails weren’t cached correctly. Beforehand, there was a difficulty the place the photographs of uploaded app thumbnails on the App Overview web page and on the drop-down sidebar record, which offers fast entry to your out there apps, sometimes appeared as clean or lacking. We mounted the problem.
Login/Logout and Signup Circulate
Launched modal pop-ups for login and signup, and for varied eventualities the place redirects have been beforehand employed
- We’ve retained devoted login and signup pages whereas introducing modal home windows for these actions. Now you can entry login and signup capabilities instantly out of your energetic web page, offering fast and quick access to those important options.
- For operations like “Use Mannequin,” “Use Workflow,” “Set up Modules,” and extra, we have now changed the redirection course of with modal pop-ups. This eliminates the additional step and ensures a smoother consumer journey.
Bug Fixes
- Fastened a difficulty with the reCAPTCHA system. We improved the reCAPTCHA system to supply customers a notably smoother and extra user-friendly expertise.
Labeling Duties
Bug Fixes
- Fastened a difficulty the place it was not potential to submit labeled inputs for some labeling duties. Beforehand, you could possibly not submit labeled inputs for sure labeled duties. When engaged on some labeling duties, for those who loaded the enter, chosen the related ideas, and tried to submit your labeled enter, you could possibly encounter an error. We mounted the problem.
Workflows
Bug Fixes
- Fastened a difficulty with processing movies utilizing the Common workflow. Now you can confidently course of movies with the Common workflow with out encountering any hindrances or points.
- Fastened a difficulty with modifying a workflow. Beforehand, whereas modifying any workflow, the mannequin model displayed “No outcomes discovered,” which was inconsistent with the preliminary workflow creation expertise. The mannequin model habits now matches what’s displayed when initially making a workflow.
- Fastened a difficulty the place it was not potential to repeat a workflow with out first altering its ID. Beforehand, you could possibly efficiently copy a workflow solely after altering the copied workflow ID. Now you can copy an present workflow, even for those who maintain the identical workflow identify in the course of the copying course of, equivalent to from “(workflow identify)-copy” to “(workflow identify)-copy.”
- Eliminated the default/base workflow from “Use Mannequin” modal. To make use of a mannequin in a workflow, go to the mannequin’s viewer web page, click on the “Use Mannequin” button on the upper-right nook of the web page, choose the “Use in a Workflow” tab, and choose a vacation spot app and its base workflow. You’ll be redirected to the workflow editor web page.
- Beforehand, for those who tried to replace the workflow within the editor web page, you could possibly encounter an error. The difficulty arises as a result of the app’s default/base workflow can’t be edited, however this limitation is communicated to the consumer late solely after they’ve made adjustments to the workflow. We mounted the problem by graying out or excluding the bottom workflow choice when customers try to make use of a mannequin in an present base workflow.
Markdown Search
Added skill to carry out text-based searches throughout the markdown notes of apps, fashions, workflows, modules, and datasets
- As an illustration, whenever you go to https://clarifai.com/discover/fashions and enter “detecting the presence” within the search area, the outcomes will embrace any related useful resource the place the phrase “detecting the presence” is discovered throughout the markdown notes. This function enhances search performance and permits for extra exact and context-aware discovery of sources primarily based on the content material inside their markdown notes.
Cowl Picture
Added cowl picture help for varied forms of sources, together with apps, datasets, fashions, workflows, and modules
- This function provides a visually partaking factor to your sources, making them extra interesting and informative. You too can delete the picture if you wish to.