Skip to main content
Before you purchase a dataset, run try_dataset on every serious candidate. The tool opens a closed trial environment and returns preview content, schema coverage, and a full quality datacard — everything you need to make a confident buy decision without spending a cent or exposing paid content.

​
Call try_dataset

Pass the dataset ID from discovery and, optionally, the exact column names your pipeline requires. Sella will report which of those fields are present in the schema, so a mismatch is obvious before you commit.
Always pass required_fields. Without it, you’ll need to scan the schema manually. With it, a missing column surfaces as a clear flag in the trial response — no guesswork.

​
Read the datacard

Every Sella listing carries a datacard — a structured quality record generated by automated benchmarks and, where applicable, human review. Work through each field before deciding to buy.
The top-level verdict from Sella’s automated quality pipeline:
  • passed — all automated benchmarks are green. Safe to purchase, subject to task fit.
  • pending — benchmarks have not run yet. Evaluate the preview rows and review scores manually before buying.
  • manual_review — borderline result; the listing is live but flagged for a human moderator. Proceed with extra caution.
  • failed — do not buy. The dataset failed objective quality checks.
The date span the data covers, for example:
Compare against your requirement before purchasing. If the range does not cover your target period, skip this dataset and return to discovery.
Per-task fitness scores assessed by Sella’s review pipeline. Use the score for your specific task, not the headline number:
A high headline average can hide a low score for the task you actually care about.
A value from 0.0 (fully human-generated) to 1.0 (fully synthetic). Training on synthetic data can introduce systematic bias and distribution shift. A cautious agent auto-rejects datasets with a score above 0.5 unless the use case explicitly calls for synthetic data.
Whether the dataset’s token count suits a target model size — particularly relevant if you plan to run a Cradle fine-tuning job:
An insufficient verdict means you will likely underfit the model. Consider combining multiple datasets before launching a Cradle job.
Trust the preview over the scores. Automated metrics can be gamed; trial.previewContent is the ground truth. If the sample rows look noisy, mislabeled, or structurally inconsistent, trust your eyes over a green qualityStatus.

​
Decision checklist

Work through these steps in order. A dataset that clears all four is ready to purchase.
1

Schema fits

Every field in your required_fields list is present in the trial schema response. A single missing column is grounds to skip.
2

Quality is green

qualityStatus is passed, or the preview rows are convincingly clean and the scores are strong enough to compensate for a pending status.
3

Task fit is high

The taskFitnessVector score for your specific task clears your minimum threshold. Ignore scores for tasks you are not running.
4

Temporal range covers your period

temporalRange spans the date window your pipeline requires. Gaps or a range that ends too early disqualify the dataset.
A dataset that clears all four steps is ready to buy. Head to Purchase to retrieve full content and settle payment.