Release Notes

Version 0.3.8 (24 February 2025)

Cobalt now fully supports Python 3.13 and NumPy 2.1.

Features

Workspace.find_failure_groups() now has a max_size parameter which limits the size of failure groups returned. Groups larger than this size will be split into subgroups.

Workspace.get_groups() and Workspace.saved_groups now return GroupCollection objects.

Quality of Life Improvements

Models can now be referenced by name in calls to methods like Workspace.find_failure_groups() or when indexing into CobaltDataset.models.

Performance metrics for all models are now automatically populated in the coloring menu, without the need to call CobaltDataset.compute_model_performance_metrics().

Fixes

A number of issues affecting data table filters were resolved.

The graph layout now updates appropriately when the widget is expanded.

Some nonpublic functionality was removed from the documentation.

Version 0.3.7 (16 January 2025)

Features

The interface of CobaltDataset.add_model() was improved to allow more customization of model metadata.

GroupCollection was added as a new API for interacting with collections of groups. It can currently be used in conjunction with the Workspace.get_graph_level() method, but over the next few releases this will be integrated with the rest of the Workspace API and provide a unified way of exploring groups discovered by Cobalt.

Performance Improvements

Keywords for groups are now computed more quickly.

A regression in the performance of Workspace.find_clusters() was resolved.

Fixes

A number of issues affecting data table filters were resolved.

The data table now updates only once when a new selection is made.

Version 0.3.6 (12 December 2024)

Features

The UI layout is significantly improved, with better use of space on wide displays.

Datetime columns in dataframes are now automatically added to the coloring menu.

Fixes

Images can now be displayed in the data table in Google Colab notebooks.

Logs from the server used to display images in the data table no longer appear in notebook output cells.

Version 0.3.5 (19 November 2024)

Beginning with version 0.3.5, Cobalt is available to install directly from the PyPI repository with pip install cobalt-ai. Users can also register for a trial or noncommercial license directly from the Python interface by running cobalt.register_license().

Fixes

A bug in the calculation of t-statistics for the group comparison table was fixed.

Opening the UI for a dataset without any embedding arrays will no longer result in a difficult to understand error.

Creating a dataset with columns of non-hashable dtype no longer results in an exception being raised.

Node labels no longer remain displayed after a node is deselected in the graph.

Version 0.3.4 (25 October 2024)

Features

Added an experimental interface for automated group analysis in cobalt.lab.

CobaltDataset objects can now be saved to disk and reloaded.

Added support for displaying images hosted on a remote server in the data table.

Paragraph breaks are now rendered in text displayed in the data table.

Text keywords now include bigrams as well as unigrams for increased specificity.

Fixes

Dialog boxes now display correctly in Visual Studio Code notebooks.

The UI is now more legible in notebooks using a dark-themed interface.

The number of rows shown in the data table is now preserved when the selected data subset changes.

Version 0.3.3 (20 September 2024)

Features

Added an option to repeat colors from the colormap when coloring by a categorical column. This can be useful in situations where there are many more categories than available colors.

Added an option to color a graph by a set of clusters created with the Workspace.find_clusters() method.

Added cobalt.setup_license() to streamline license key configuration.

The Cobalt UI now supports Google Colab notebooks.

Performance Improvements

Improved responsiveness of the coarseness and connectivity sliders.

Fixes

Fixed a problem where some categorical columns could be detected as text and automatically used as input to keyword analysis.

Fixed a problem where trying to view a graph with too many nodes could cause the UI to crash.