CoreWeave Notebooks brings reactive Python analysis into Forge projects

From dashboards to custom analysis

According to CoreWeave, dashboards answer many questions about a model, such as the loss curve, the eval score or a comparison across a sweep. The questions that decide whether a model ships usually come after that: why a data slice regressed, which examples fail and what they share, and whether a pattern is causal or an artifact of how the data was collected.

Answering them requires code to slice and visualize the data already on screen. A notebook opens from a Forge project already connected to its experiments, artifacts and evaluation results. Users can query training metrics across experiments, load an evaluation table from an artifact, filter it to the relevant examples and start analyzing in the first cell, the company says.

A driving model example

CoreWeave illustrates the idea with a demo notebook. A driving model is trained to stop at red lights, and its aggregate traffic-light loss looks fine. The training data, however, has attributes the dashboard does not chart by default: weather conditions, time of day, and whether another red object is in the frame near the light.

In the notebook, the loss turned out to be low in most weather conditions but much higher in heavy rain at night when a red object is near a traffic light. Sorting frames by loss and viewing the worst ones showed traffic lights in night rain with a red building or truck nearby. According to CoreWeave, the model separates the light from these objects in daylight but confuses them in these conditions.

The company's point is that overall metrics can hide failures like this. Finding them means filtering by specific attributes, plotting comparisons an existing dashboard may not support, and inspecting the worst cases.

Reactive execution and sharing

marimo builds a dependency graph from variable references. When a variable changes, dependent cells run automatically, or can be marked stale for expensive computations. Deleting a cell removes its variables from program memory. CoreWeave says this keeps code, program state and visible outputs aligned, so an old output does not survive beneath new code during debugging.

It also makes exploration interactive. With a slider, dropdown or table filter, the notebook responds as the control moves. In the driving example, a weather selector re-renders the loss chart and the worst-frames grid together.

Notebooks are stored in the project rather than in local files. Each project has a listing page showing the team's notebooks, and Forge has an overview page that lists notebooks across projects, with search and filtering. A colleague can open a notebook with the data attached, read the investigation and copy it. When new experiments arrive, the same notebook can be rerun on the new data, which CoreWeave says turns a one-time investigation into a team regression check.

Where it fits in the AI loop, and what comes next

CoreWeave places Notebooks across the development loop: exploratory data analysis before training to check for gaps and duplicates, diagnosing an experiment that diverged from its baseline mid-training, and finding failure modes that an aggregate score averages away at evaluation.

The public preview is built for interactive analysis next to the data. Over the coming months, CoreWeave plans several extensions:

• Publishing a view from a notebook back to a Workspace as a persistent panel that refreshes with the latest data.
• Hosting a notebook as a data app, an interactive dashboard that teammates can use without opening the code, such as a labeling tool or a sliceable view of results.
• More compute in the notebook sandbox, so data science teams can train forecasting and classification models in a notebook.
• Co-authoring with CoreWeave ARIA, which will soon be able to write notebook cells from a described view, so users need not be fluent in pandas and plotting libraries.

What it means for teams

For teams already working in Forge, the change is that custom analysis no longer has to happen in separate local tools. Notebooks can be opened from any project, or started from the getting started page, and analysis stays attached to the data and visible to colleagues. The planned panel and data app features suggest CoreWeave intends results from notebooks to feed back into the Workspace, though those are described as upcoming work.