Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.chemolytic.com/llms.txt

Use this file to discover all available pages before exploring further.

A registered model is a trained model snapshot promoted from an experiment trial. Once registered, a model is permanent: it can be deployed, downloaded, and versioned, independent of the experiment that produced it.

​
Concept

When you click Register Model on a successful trial, three things happen:
  1. The trial’s pipeline (preprocessing + algorithm + hyperparameters) is captured
  2. The model is refit on the full training set (no CV split) for production-quality predictions
  3. A version is created, pinned to the dataset that was used for training
This means models are reproducible, auditable, and stable. Even if the source experiment is later deleted or the dataset version changes, the model retains everything it needs to make predictions.

​
Models page

Go to Models in the project sidebar.
Models list page showing model name with version badge, algorithm, target, dataset, status, and CV score
ColumnDescription
NameModel name and version badge (e.g., Brix Predictor v3)
ModelAlgorithm type (PLS, Ridge, RF, etc.)
TargetProperty the model predicts
DatasetPinned dataset used for training
StatusBuilding (pulsing), Ready, or Failed
CV scorePrimary metric (RMSE for regression, F1 macro for classification)
CreatedDate registered

​
Tabs

TabShows
ActiveModels in normal use (default)
ArchivedArchived models, hidden from main list
AllBoth
The page polls every 3 seconds while any model has status Building so progress updates live.

​
Status flow

StatusMeaning
BuildingModel is being refit on the full training set in the background
ReadyRefit complete, model is deployable
FailedRefit failed; see the error on the detail page
Building usually takes 10 seconds to a few minutes depending on dataset size and algorithm.

​
Model detail

Click any model in the list to open its detail page.
Model detail page showing performance metrics, pipeline, and version history
Shows the model name, version badge, status, description, and metadata: algorithm, target property, source dataset, and source experiment/trial (clickable links to the original).

​
Performance hero

A metric grid showing the model’s performance, with a CV/Test toggle. The metrics shown are the same as in the trial detail (R², RMSE, MAE, Bias, RPD for regression; Accuracy, F1, Precision, Recall for classification). See Trial results for how to read each. These numbers come from the original trial’s CV and test split, not from the refit. The refit just trains on more data; the metrics reported here represent how the model is expected to perform on new spectra.

​
Tabs

TabContents
ResultsPipeline visualization, predicted vs actual chart (regression), or confusion matrix (classification)
DetailsExact hyperparameters, preprocessing parameters, and traceability info (source experiment, trial, artifact path)
VersionsEvery version of this model, sortable, with current version highlighted

​
Downloading a model

If your plan allows it, click Download to get a .joblib file containing the trained model.
Model download requires the allow_model_download plan feature. If your plan doesn’t include it, the Download button shows a gem icon and clicking it opens an upgrade dialog instead.
The file is named {model_name}_v{version}.joblib and contains the full trained scikit-learn pipeline. You can load it with: 
import joblib
model = joblib.load("Brix_Predictor_v3.joblib")
predictions = model.predict(X)  # X is shape (n_samples, n_features)
This is useful for:
  • Offline predictions in air-gapped environments
  • Integrating into custom Python workflows
  • Auditing the model architecture
A dedicated Python SDK is planned. It will package the model with full metadata (sensor, dataset version, expected x-axis, target property, units) so you don’t have to track these separately. For now, the raw .joblib file contains only the fitted scikit-learn pipeline.
The downloaded model expects spectra with the exact same x-axis as the sensor used to train it. Predictions on spectra from a different sensor will fail or produce nonsense.

​
Lineage and reproducibility

Every registered model includes:
  • A link to the source experiment (or “Deleted” if removed)
  • A link to the source trial (or “Deleted” if removed)
  • The pinned dataset it was trained on (immutable)
  • The exact pipeline and hyperparameters
This lineage ensures you can always trace any prediction back to the data and configuration that produced the model. Even if the experiment is deleted later, the model itself stays intact.

​
Editing a model

You can rename a model and update its description from the detail page. The pipeline, dataset, and metrics are immutable.

​
Deleting a model

You cannot delete a model that has active deployments. Remove its deployments first. The error: “Cannot delete model with active deployments. Remove all deployments first.”
Click the trash icon on the detail page, confirm in the dialog, and the model and its artifacts are permanently removed.

​
Plan limits

Your plan limits how many registered models you can have per project (max_registered_models). The current count and limit appear at the top of the Models page.