What kinds of datasets can you manage and annotate on DagsHub?

DagsHub handles multimodal datasets — vision, audio, and LLM data — on one platform. You can connect multiple data sources, then enrich, query, visualize, and annotate them, with the goal of turning large volumes of raw data into "golden datasets" for model improvement. The practical limits depend on your plan: annotation for public repositories is available on the free Individual tier, while multimodal annotation and auto-labeling are Team features.

What data types are covered

DagsHub describes its dataset work as multimodal, naming three families explicitly:

• Vision — image data for curation and annotation.
• Audio — audio data, handled through the same curation and annotation workflow.
• LLM data — text data used for language model work.

The platform positions this as transforming "petabytes of vision, audio, and LLM data into golden datasets." So the answer to "can I put my data type here" is broad: if it falls into image, audio, or text/LLM categories, it's in scope.

What you can actually do with a dataset

The workflow described on the site has four verbs attached to it: connect, enrich, query, visualize, annotate.

• Connect multiple data sources. You are not limited to one storage location; the platform is built to pull from several sources.
• Enrich and query. Datasets can be enriched and queried rather than just stored.
• Visualize and annotate. Annotation happens in a dedicated annotations workspace.
• Version and track lineage. Data versioning and lineage are listed as included capabilities, which matters if you need to know which data produced which model.

Alongside datasets, the same platform covers experiment tracking (MLflow-compatible) and model management with full lineage from model back to source data — useful if your dataset work feeds directly into training runs.

Annotation capabilities by plan

This is where the dataset question gets a concrete answer, because annotation features are tiered.

Two things worth reading carefully:

1. On the free Individual tier, the annotations workspace is for public repositories. If your dataset is private and you need to annotate it, that points to Team.
2. Multimodal annotation and auto-labeling are Team features. If your work is audio or vision annotation at any volume, the free tier's annotation scope is the constraint to check first.

Team also lists Label Studio compatibility, which matters if you already have Label Studio projects or labeling conventions — it suggests your existing annotation setup can plug in rather than being rebuilt.

Choosing based on your dataset situation

• Public, small, mostly text or experimentation: the Individual tier covers unlimited public repositories, unlimited experiment tracking for public repos, and up to 100 tracked experiments in private repos, with 20GB of storage. Annotation is available for public repos.
• Private multimodal data you need to label: Team, because multimodal annotation, auto-labeling, connecting your own storage, and Label Studio compatibility all sit there. Team also raises the ceiling to 1TB or 2 million files and starts with a 14-day free trial.
• Petabyte-scale or on-premise requirements: Enterprise, which adds petabyte-scale data management, deployment to your own cluster, VPC/air-gapped installation, SSO/LDAP/OIDC, and OpenShift compatibility.

A concrete way to test fit

Pick one real dataset you'd want to annotate — for example, a few hours of audio you need transcribed and labeled, or an image set you need bounding boxes on. Then check two things against the table above: whether that repository can be public, and whether the annotation type you need is listed for your tier. If either answer is no, the Team tier is the level to evaluate, and the 14-day trial lets you run the actual annotation workflow before committing.

The site reports over 65,000 data scientists building AI with DagsHub, and lists integrations with common ML frameworks, open-source formats, cloud storage, and MLOps tools — so the dataset layer is designed to sit inside an existing stack rather than replace it.