Simple pricing for AI teams and solo builders.

Dataset Finder is a workspace for AI training data. You search a curated catalogue of tens of thousands of machine learning datasets in plain language, read a full dataset card before committing, organise what you find against your AI projects, and commission custom data annotation when nothing suitable exists. It replaces the scattered mix of directory browsing, paper reading, spreadsheets and labeling vendors most teams juggle today.

No. Dataset Finder does not host, store, or redistribute datasets. We are a curated catalogue and a workspace built on top of data that already exists in the world, not a file repository.

What we maintain is the metadata and the intelligence around each dataset: where it lives, who published it, its licence, size and format, quality scores, biases, limitations, and suitable use cases. When you bookmark a dataset, your workspace stores that record and your notes, never the files. Dataset Finder sends you to the origin, whether that's Hugging Face, an academic page, a government portal or your own storage, where you access it under its own licence terms.

The one exception is custom data: when you commission an annotation project, our team produces and delivers that dataset to you directly.

Yes, because we never take possession of or resell anyone's data. You obtain each dataset from its original publisher, under that publisher's licence. To make that easier, every dataset card carries a Licence field and a Licence check, so you can see whether a dataset is permissive, non-commercial or research-only before you invest time in it. We surface the information; you keep the relationship with the source. Always review the licence yourself before training on a dataset.

Those are repositories: enormous, keyword-driven and largely unevaluated. You still judge quality yourself, usually by reading papers and sampling files. Dataset Finder searches across those sources and adds the layer they don't have: human curation, quality and bias analysis, label-density and class-balance signals, suitability scoring, and a workspace to track decisions. And when nothing in any repository fits, you can order the annotated data instead of giving up.

It means a human looked at it. Datasets are reviewed and enriched by our team rather than crawled in bulk, and we refresh the catalogue continuously: new datasets, corrected metadata, updated licences, dead and moved links swept out. That's the difference between a search index and a catalogue you can trust.

Yes. Create a workspace per team, client or product line and invite colleagues into it. Bookmarked datasets, statuses, notes, AI projects and cloned recipes are shared, so two people don't evaluate the same dataset twice and the work survives someone changing role.

Order it. Describe the data you need, attach your annotation guidelines and a few labeled examples, and our team handles data collection, labeling, quality assurance and delivery. You can self-serve with credits for typical projects up to a few hundred thousand items, or talk to our team for large-scale work with custom SLAs, NDAs or on-premise delivery.

The full range of supervised labeling work, across image, video, text, audio, multimodal and document data:

Computer vision. Bounding boxes, polygons, semantic and instance segmentation masks, keypoints and landmarks, object tracking across video frames, image classification and tagging.
NLP & documents. Named entity recognition (NER), text classification, sentiment and intent labeling, relation extraction, OCR correction, document layout and key-value extraction, translation and multilingual annotation.
Gen-AI & LLMs. Instruction and prompt/response pairs for supervised fine-tuning (SFT), preference ranking and RLHF, red-teaming, content moderation, factuality and hallucination review.
Audio. Transcription, speaker diarization, event and intent labeling.

We also cover regulated domains such as medical imaging and clinical text, where annotation is carried out by domain experts rather than generalists.

Yes, and this is usually the preferred setup. Our annotators can log into your annotation platform and work in your environment, your schema and your review workflow, so the labeled data lands where your pipeline already expects it and never leaves your tenancy. We're experienced across the common tools, including Label Studio, CVAT, V7, Labelbox, Encord, Roboflow, SuperAnnotate and Prodigy, and can work in an in-house or proprietary tool too.

If you don't have a platform, we'll provide and administer a secure one for your project, so tooling never becomes the blocker.

Fundamentally different model. Crowdsourcing marketplaces distribute micro-tasks to an anonymous, rotating pool paid per click, which is why they struggle with complex guidelines, domain knowledge, consistency over time and any kind of traceability.

We don't use a crowd. Annotation is done by in-house teams who are recruited, trained and fairly employed by Innovatiana, working from our own delivery centre, supervised by a dedicated Data Labeling Manager who is your point of contact. The same people stay on your project, so guidelines compound instead of resetting, and every annotation is traceable to a known annotator. For specialist work we staff domain experts such as medical students and clinicians, linguists, lawyers and developers, rather than generalists.

Practically, that means higher inter-annotator agreement, workable complex taxonomies, and an audit trail you can show a regulator. It's also why we can commit to ethical sourcing rather than opaque subcontracting.

Quality assurance is built into the workflow, not bolted on at the end. We start from your guidelines and a calibration round on a sample, then run manual sampling reviews, inter-annotator agreement (IAA) measurement, gold-standard comparison and automated consistency checks throughout the project. A dedicated manager monitors throughput and label consistency, and edge cases are escalated back to you rather than guessed. You get corrected, production-grade ground truth, not a first pass.

We deliver in the format your training pipeline expects: COCO, YOLO, Pascal VOC, JSON, JSONL, CSV, XML, or your own custom schema, transferred securely or written straight into your storage. On timing, we assess your needs and run a test batch where useful within about 48 hours, so you can validate quality on real examples before committing to full volume. Delivery timelines then depend on volume and annotation complexity, and are agreed up front.

Yes. We assess the sensitivity of whatever you entrust to us and apply the matching controls: NDAs, restricted access, secure transfer, and working inside your environment where required. Our practice is built around GDPR and EU AI Act readiness, with full traceability of who annotated what, which matters when you have to document your training data provenance. Data handling terms are agreed before a project starts.

There is, and it needs no credit card. Free gives you dataset discovery across the whole catalogue, dataset card essentials and one demo project. Startup (€9.90/mo) unlocks full dataset cards, unlimited saved datasets and multiple projects. Pro (€29.90/mo) adds unlimited projects, the full recipes library, exports and compliance reports, and training data lineage. Custom annotation is priced separately with credits, charged per task or per dataset delivered with no subscription or set-up fee, and paid plans carry a discount on those credits.

Dataset Finder is built by Innovatiana, a data labeling company with in-house teams of professional annotators and AI trainers who have delivered training data to 200+ AI builders since 2021. Not a crowdsourcing marketplace: fairly employed annotators, full traceability, and GDPR and EU AI Act readiness by default.