Searching datasets

Dataset Finder uses semantic search, not keyword matching. That means you describe your intent — the AI product or model you want to build — and the engine understands meaning, not just the words you typed.

You can search in any style:

• A vague idea or product concept
• A precise technical specification
• An AI use-case description
• An exact dataset name or keyword (like a known benchmark)

Short or long, technical or plain English, the system reads for intent and returns the datasets that actually fit your need.

For complex needs, open the detailed search request window. Press Ctrl + K (or select the expand icon next to the search bar) to open it. It gives you a larger text area to write a long, natural-language query with multiple criteria at once.

You'll reach a character counter (for example 0/500) showing how much room you have. The limit depends on your plan:

A longer allowance lets you stack context, for example: "Find open-source CCTV datasets from hotels in Europe with at least 1M frames and labeled entrances." Type your request, then select Run search.

Every search returns up to 33 results, organized in two groups:

• Top picks — the 3 strongest matches for your query, shown first as detailed cards.
• More suggestions — the next 30 results in a scannable list with type, name, volume, file format, and match score.

Each result card shows a small picture, a type icon (image, video, audio, text, medical, or multimodal), a short description, and its relevance score. Select any result to open its full dataset card in the panel on the right.

Every result carries a relevance score from 0 to 100 in a small bubble. It shows at a glance how well that dataset matches your specific query.

The bubble is colour-coded so you can read it instantly: the closer to 100, the greener it is; the closer to 0, the redder it is.

The colour lets you scan a list of results and spot the strongest candidates without reading every score, then open the green ones first.