Release notes

Major updates

• Metric changes — In this release, we wanted to increase the versatility of how evaluation metrics that are used in Benchmark are behaving. Hence, we are adding more details to Metric cards to be more descriptive. Adding details like the input, output, pricing, and supplier of each Metric. This would also give more visibility to the pricing of Benchmark jobs.
• Text normalization setting in Benchmark — A new feature has been added to Benchmark which allows using various normalization settings for text to be used. This feature comes with minor UI improvements to Benchmark job configuration.

Minor updates

• Added the ability to use public URLs as valid data in Dataset and Corpus upload.
• Added the ability to filter models in Discover which are FineTune compatible.
• Added tooltips for different types of nodes in Design when you hover over them.
• Added double-click functionality in Design to re-align zoom and access node details.
• Improved UI visuals of Design nodes with unified asset node types and dynamic font size.
• Added autosave functionality for pipelines in Design.
• Added a Dashboard guide that is visible for the first time a member signs.

SDK updates

• SDK for FineTune — SDK now supports fine-tuning models.

Bug fixes

• Fixed an issue with data persistence that kept experience progress when switching teams.
• Fixed an issue with handling third party packages which impacted certain nodes in Design.

Major updates

• Corpus and Dataset changes — In release 2.2.0 we introduced a new asset type “Corpus” which allowed structured data to be onboarded on aiXplain without having to allocate a function to it.
  • Added the ability to create Datasets from Corpora, or use them directly with your favorite tools like Benchmark and FineTune.
  • Improved Dataset creation process to allow for more options such as adding multiple references and handling errors.
  • Added Datasets and Corpora creation from files already uploaded to aiXplain instead of only supporting local files.
  • Added Dataset and Corpus logs which are available on the details page of the assets to show the history of these assets.
• AutoMode training temporarily disabled — In light of the changes to datasets in aiXplain, our team is working hard to address their compatibility with AutoMode. This won’t affect the already trained AutoMode models. Expect an update that will address this soon.

Minor updates

• Added more information on the Corpus details page that shows the number of columns/features and rows/segments in a Corpus.
• Added credit expiration notice for Credits granted through phone verification.
• Updated the colors in SHAP plots inside Benchmark reports.
• Added more metadata categories in Benchmark reports.

Bug fixes

• Fixed a bug that showed the “Compare” button in the asset drawer for assets that don’t support this function.
• Fixed a bug that displayed the wrong labels in nodes inside Design.
• Fixed an issue with certain models that failed when rerunning a Benchmark job.

Major updates

• Discover changes — In preparation for bigger upcoming changes to some of our AI asset types, we updated the design of Discover to make finding the assets that you are looking for easier and more intuitive! The changes will also be reflected on your assets that are accessible from your Dashboard.
• Introducing a new asset type: Corpus — In this release we added a new asset type that you can find, onboard, and use in aiXplain. The corpus is structured data which is not defined to be used for a single particular AI function. There are limited things that you can do with corpora for the time being but our next release will allow extracting multiple datasets for multiple AI functions from a single corpus. For example, I can onboard a corpus with columns for speech, transcript, translation, speaker age, speaker gender, speaker ID where I can use it to create datasets for Speech Recognition, Speaker Diarization, Speech Classification, and Translation.
• Updates to API access keys — We previously added the ability for teams to generate access keys from the team settings. In this release, we removed the auto-generated default access key which was specific to a particular model or pipeline for each team. This change allows creating and using different access keys for different endpoints. With added security allowing members to delete and recreate access keys as they see fit.

Minor updates

• Translation metrics added in Design — We added all of the Translation metrics found in Benchmark to the metric node inside Design.
• Improved Benchmark — We improved the way that scoring in Benchmark takes place. Those changes will have positive impact on Benchmark performance and stability.
• Enhanced network security — We added improved encryption to the tokens sent out over the network.
• User guide in Dashboard — Added a user guide in the Dashboard to help the members access the documents page and use the platform.

Bug fixes

• Fixed a bug that made Benchmark reports appear to be stuck at 90% when they are complete.
• Fixed a bug where members couldn’t find the “switch team” option when the team invitation is accepted.
• Fixed a bug that prevented the phone number verification to appear on the dashboard.
• Fixed a bug in displaying metrics inside the metric node in Design that collapsed the scroll.
• Fixed a bug where some function names were displayed incorrectly.

Major updates

• FineTune tool — We have launched the ability to fine-tune select models for MT and ASR. You can also view the logs of the model inside the details page for it. Read more
• Segmentor node and constructor node — We added new nodes into Design! Using the segmentor node and constructor node, members now have the ability to customize the segmentation of the data flowing through their pipelines and reconstruct them back. This optimizes the use of Design for models where the input size matters.
• Metric node — We added another new node in Design! Members can add benchmarking to their pipelines as a node which gives them the ability to evaluate the performance of their models while a pipeline is running.
• Activity logs — We added a log for all your activities on aiXplain which allows you to filter them by product and by status. This will give members more control and insight over what activities are being done on their teams such as model fine-tuning and Benchmark reports.

Minor updates

• Imputation of failed inputs — Now in Benchmark, members have the option to impute the scores of some failed segments to get a higher representation of their data.
• Added the number of segments to be displayed onto the asset card for datasets.
• Added the ability to create API keys on aiXplain: In the top menu under team settings, members will be able to generate API keys which can then be used as a universal key to run all of their assets.
• Removed phone number from registration.
• Added icons next to AutoMode and FineTune models.
• Added multiple notifications in platform for Benchmark, FineTune, and AutoMode.

Bug fixes

• Fixed an issue which didn’t allow saving pipelines without setting decision nodes.
• Fixed a bug that showed a redundant message when not having enough credits on the platform.
• Fixed a bug that allowed users to add multiple datasets in benchmarking when they shouldn’t.
• Fixed a bug where submetrics showed redundant metrics name.
  Fixed an issue with the IQR filter displaying inside the segments tab in a benchmark report.
• Fixed a bug where the label of the decision node couldn’t be updated.
• Improved error handling for model and pipeline tryout.

Major updates

• Benchmarking for text classification models — We have expanded our benchmarking capabilities to support text classification models. The current supported functions are Sentiment Analysis and Offensive Language Identification.
• Subtitling node — We added a new node in aiXplain’s designer, the subtitling node.
• AutoMode for ASR — Following AutoMode for MT, you can now create your own AutoMode ASR model from Discover.

Minor updates

• Bring your own models output — We have added the ability for users to upload their models output during the dataset upload and compute the score of that model without the model being onboarded on aiXplain.
• Bring your own bias categories — Similar to bringing your own model’s output, during dataset upload you can identify categorical columns that you can use for bias analysis and topic classification in benchmarking.