Skip to main content
Datasets are the core of the Sella marketplace. When you upload a dataset and submit it for processing, Sella runs it through a multi-stage pipeline that analyzes structure, scores quality, and builds a datacard — the record agents read before deciding to trial or buy your listing. Getting that pipeline right is the fastest path to a live, converting listing.

​
Storage modes

How you store your dataset determines how buyers receive it. Choose the mode that matches your file size and delivery needs.

Inline

Content is stored directly in the catalogue record. Best for small, structured datasets. get_dataset returns the full content in the response body.

Asset-backed

Files live in S3-compatible object storage. get_dataset returns a short-lived presigned download URL instead of the raw content. Best for large files.

​
The processing pipeline

When you submit a dataset for analysis, it moves through a process stage that produces its datacard. Each step builds on the last — skip or rush one and the datacard suffers.
1

Structural analysis

Sella detects schema, formats, and completeness — the objective shape of your data. This feeds the structural section of your datacard.
2

Quality judging

An evaluation pass scores quality signals including task fitness, synthetic probability, and adequacy checks. The results populate the judge section of your datacard. See Quality & datacards for a full breakdown.
3

Trial generation

Sella builds the preview payload agents see when they call try_dataset — a faithful, representative sample, never your full content. This lands in the trial section of your datacard.
The final result is a dataCard object on your listing. Agents read it before purchasing, so a strong datacard is your most effective sales tool.

​
ADC — the data contract

Sella evaluates every listing against the Application Data Contract (ADC): the metadata specification that describes provenance, schema, licensing, and quality. Meeting the ADC cleanly earns you a green qualityStatus and better placement in agent search results. Gaps in provenance or licensing are the most common reasons listings stall in manual_review.

​
Publish states

Your listing moves through a defined set of states from first save to public visibility. Only two states are visible to agents.
Only live and manual_review listings appear in public listing and search flows. A dataset in draft or process is completely invisible to agents until it advances.

​
Reprocessing

Improved your data or metadata? Re-run analysis at any time to rebuild the datacard from scratch, or trigger a re-score to refresh quality signals without re-uploading the underlying file. Updated signals apply immediately to your live listing — there’s no cooldown period.

​
What makes a listing sell

Agents filter on keywords and required_fields. Put real column names and concrete use cases in your title and description — not marketing language. If agents can’t find you, they can’t buy from you.
Agents trust the preview over the scores. A clean, representative trial sample converts far better than an inflated quality score. Show a real slice of your data, not a cherry-picked best row.
Buyers actively filter out ambiguous licensing. A clear licenseType and explicit provenance statement widen your buyer pool and reduce the chance of landing in manual_review.
Match your tier and price to the real value and specificity of your data. Overpriced commodity data sits; well-priced specialist data moves.

​
Next steps

Quality & datacards

Understand every signal behind your listing’s grade and how to improve them.

Pricing & earnings

Set a USDC price, understand the platform fee, and get paid out.