New Features

New Functionality for Monitoring Large Language Models:

Users can now onboard generative text models and monitor sequence lengths and token likelihoods, providing deeper insights into language model performance

Users can create “stage” tags for models using existing model tagging functionality

Easily display stage tags on the model Home Screen for improved organization and management

New Functionality for CV & NLP Models: Users can now use raw anomaly score as the default shown anomaly score column and sort inferences by raw anomaly score, for easier analysis of anomalies

New Feature: We've introduced an optional configuration feature that enables users to access the Admin Console using a different IngressClass, providing greater flexibility and security

Enhancements

Global Filter by Inference and PartnerID

In global filters, users now have the ability to filter by inference and partner inference ID

Sort Inferences by all columns

Updated Functionality: Users can easily analyze their data by allowing them to sort inferences by all columns (except for inference_ID) on the inferences deep dive page

Increase compute resource allocation by default to anomaly detection

We've increased compute resource allocation by default to anomaly detection enrichments in high-availability environments and allow for user override as needed for customization

Bug Fixes

• We've fixed a bug that caused issues when training models with attributes that have lots of different values. Now, any attribute with more than 50 unique values will be filtered from the dataset before training so you can get more accurate results
• We've resolved a bug that caused OIDC users to see a generic message on their home screen instead of their first name
• Resolved an issue related to StatefulSet backups, ensuring that AWS VolumeSnapshots are correctly initiated by the backup procedure
• Resolved an issue that previously prevented OIDC users from accessing the full functionality of global filtering for individual model screens
• Resolved a bug that caused global filters to inaccurately round down timestamps to the previous day, resulting in loading issues
• Resolved the problem where JSON file uploads were not immediately registered upon API call, causing delays in batch appearance until after the ingestion process was completed