Five Things Every Dataset on LabelSets Now Ships With

Verified contamination-free against 13 ML benchmarks

The single biggest hidden risk in training data is benchmark contamination — when a dataset secretly contains samples from the eval set you're about to benchmark on. The result: inflated metrics, unreproducible results, and embarrassing post-launch discoveries.

Every dataset on LabelSets is now scanned at upload time against MinHash signatures of 13 known eval sets, plus characteristic-phrase matching as a second layer. Datasets that pass receive a SHA-256 contamination certificate and the Eval-Clean badge. Datasets that fail are flagged with the specific benchmark they overlap with and a similarity score.

Multi-layer plagiarism + republishing detection

Every upload is checked against four originality signals before it can be listed:

• MinHash LSH — 128-element near-duplicate detection across the entire LabelSets corpus, catches reformatted copies of existing datasets
• Statistical fingerprinting — column distributions, vocabulary profiles, and color histograms detect structural copies even if the content is shuffled or renamed
• External registry search — HuggingFace and Papers With Code lookups catch sellers republishing public datasets with new names
• Async web sampling — random text passages from the dataset get checked against the open web post-publish, with the seller flagged if hits cluster

Datasets scoring above 0.95 originality earn the Original badge and an originality certificate. Datasets that match an existing source are blocked from listing.

Auto-generated structured documentation following Gebru et al. (2018)

The "Datasheets for Datasets" framework was proposed by Timnit Gebru and co-authors in 2018 as the gold standard for ML dataset documentation. The catch: writing a good Gebru datasheet by hand takes 4-6 hours per dataset. Almost nobody does it.

LabelSets now generates a complete Gebru datasheet automatically for every published dataset, populated from validation stats, provenance, label types, collection method, and seller-provided metadata. Buyers see seven structured sections: motivation, composition, collection process, preprocessing, intended uses (and unsuitable uses), distribution, and maintenance.

The datasheet renders inline on every dataset page under the new Intelligence tab, and is also available as raw markdown via GET /api/datasets/:id/datasheet?format=markdown for buyers who want to include it in their model card.

Multi-framework compliance assessment with score and gap analysis

Every dataset gets a full compliance report covering the four frameworks that matter for production ML: GDPR, CCPA, HIPAA, and the EU AI Act. Each report includes a 0-100 compliance score, identified gaps, recommendations, and a list of certifications the dataset qualifies for.

The report is generated by a structured Claude prompt working from the dataset's provenance fields (collection method, consent type, license, jurisdiction) and is stored as both structured JSON and rendered HTML. Buyers can read it before purchase and download the HTML version for their compliance file after buying.

Multi-factor fair-market value estimate with confidence range

Every dataset gets a fair-market valuation computed from five weighted signals:

• Uniqueness (35%) — pulled from the originality engine
• Demand (25%) — derived from search logs and zero-result query patterns in the dataset's category
• Quality (20%) — the LabelSets Quality Score (LQS) composite
• Comparables (15%) — price positioning vs similar published datasets
• Outcomes (5%) — average improvement reported by buyers who trained on it

The result is a USD estimate with low/high confidence range, displayed both as a card on the dataset page and as a price-vs-AI-valuation delta in the seller dashboard so sellers can see when they're under- or over-priced relative to the model.