Dataset Studio

Dataset Studio is the data preparation and management surface. It supports importing, labeling, annotating, splitting, preprocessing, and versioning datasets — without rewriting the original files.

Image tasks

Dataset Studio covers computer vision datasets across image classification, object detection, and segmentation.

• Annotation tools — a dedicated annotation editor with select/move, bounding-box drawing, and polygon segmentation drawing.
• Classification — upload classification images with labels, or bulk-relabel a selection.
• Format-agnostic imports — YOLO and COCO formats are supported natively. Reference datasets stay read-only.

Text tasks

The same studio handles text records for classification, summarization, and question answering. See NLP Capabilities for the task-specific detail.

Dataset management

• Unassigned inbox — new uploads land in an unassigned inbox. Proceeding splits them into train/valid/test at a default 70/20/10 proportion.
• Exploratory data analysis — a dedicated EDA tab gives a real-time health check: split counts, class balance, geometry constraints, and warnings.
• Preprocessing and augmentation — apply allowlisted preprocessing (resize, normalize) and augmentations. The studio stores materialized copies under versioned artifacts, so original uploads are never altered.
• Versioning — freeze a materialized, preprocessed snapshot as a named dataset version, so a training run always points at an immutable input.

Importing from the Hugging Face Hub

Rather than starting empty, pull an existing dataset straight from the Hub. Browse from the catalog, choose a dataset, and Orinth imports it into the active project as a first-class dataset — ready to annotate, split, preprocess, and version like any other.

Imported datasets are grouped under Imported from HuggingFace in the catalog, so their provenance stays clear.

To generate a dataset from your own source documents instead, see Data Recipes.

Why versions matter

A training run records the dataset version it consumed, not the dataset. That is what makes a run reproducible from its record alone: the input is frozen, the config is stored with the job, and the artifacts are written under that run's directory.