# Vertex AI Dataset

> Vertex AI datasets are lightweight registry objects: the managed dataset itself adds no meaningful charge, while the underlying training data bills through Cloud Storage or BigQuery. ZopNight inventories every dataset via Cloud Asset Inventory and uses them to link training pipelines to their data sources across 4 modalities: tabular, image, text, and video.

Source: https://zop.dev/integrations/gcp/vertex-dataset
Updated: 2026-08-19

---

A Vertex AI dataset is a managed collection of training data for tabular, image, text, or video models. The dataset object itself is lightweight; the underlying data bills through Cloud Storage or BigQuery.

## A pointer, not a payload

The managed dataset is essentially a structured reference: it records where training data lives and how it is labeled across the four supported modalities of tabular, image, text, and video. The object adds no meaningful charge of its own. What bills is the substrate underneath: source files in Cloud Storage at that product's storage rates, tabular sources under BigQuery storage meters, and any labeling work performed against the data.

## Datasets as the join key of ML topology

ZopDev inventories datasets via Cloud Asset Inventory to link training pipelines with their data sources. In the ML topology ZopNight assembles, the dataset row answers the question "what data trained this model", which turns an otherwise mysterious storage bucket into an attributable training input with a discoverable owner and purpose.

## Training data that outlives its models

Dataset-shaped waste hides downstream of the object itself. Source buckets stay at full size for models that retrain quarterly from fresh data and will never reread the old corpus. Labeled datasets persist for projects that shipped or died. Duplicate exports of the same tables sit staged for experiments that ended without cleanup. The managed dataset makes each of these findable: an orphaned dataset whose last associated training pipeline ran long ago points directly at storage worth archiving or deleting.

## Managed datasets in the console

Google Cloud console → Vertex AI → Datasets lists datasets per region with modality and creation date. Cross-referencing that list against Model Registry activity is the quickest way to spot training data no pipeline has touched in months. Since the object itself is free, keeping datasets registered costs nothing. The audit value lies entirely in the links they preserve.