Datasets
Create, version, and evaluate datasets — collections of traces grouped for streamlined analysis — in Stratix Premium.
A dataset is a named, versioned collection of traces grouped for streamlined analysis and evaluation. Datasets are the reusable, reproducible input to a trace evaluation: assemble the traces once, then run judges and scorers against the set as many times as you need. Find them under Catalog → Datasets in the left rail.
What you can do
Browse every dataset in your project, with Total Datasets, Total Rows, and Versions stat cards
Search datasets by name
See each dataset's row count, column count, current version, and last-modified time
Open a dataset to preview its rows, inspect its schema, and review version history
Run an evaluation directly against a dataset
Delete a dataset (the traces it references are not deleted)
Creating a dataset
Click New Dataset for three paths:
Create from traces
You already have traces in the project — pick a filtered set and save it as a dataset.
Upload traces from file
You have traces in a file (single trace or batch) to bring in and group.
Generate synthetic traces
You want to create new trace data — from existing data, an industry scenario, or a saved definition. See Synthetic data.
No traces yet? Generate synthetic traces builds a full, realistic multi-agent dataset from a built-in industry scenario in minutes — no instrumentation required. See Synthetic data.
Inspecting a dataset
Opening a dataset shows three tabs:
Data Preview — a sample of the dataset's rows: each trace's events, status, end time, metadata, trace ID, framework, and agent name.
Schema — the columns and types the dataset conforms to (used when appending: an appended trace's schema must match).
Version History — every version of the dataset and when it was created.
From a dataset you can Add Traces (append more, deduped by trace id) or Run Evaluation to score every trace in the set.
Versioning
Datasets are versioned. Each version pins an explicit list of trace ids and a row count; the current version is what evaluations and the preview read. Today, adding traces dedupes the new trace ids into the current version in place (so you never get duplicate rows, and the row count reflects the deduped set). Full multi-version snapshots — where an append creates a new immutable version so an evaluation you ran against v1 stays reproducible after the append — are scaffolded in the model and on the near-term roadmap.
Provenance
Synthetically generated datasets are tagged with their origin (provenance_kind = synthetic) and the generation run that produced them (provenance_run_id), so you can always trace a dataset back to the job and configuration that created it.
Where to next
Synthetic data — generate trace datasets from scratch
Traces — the individual records datasets are built from
Trace evaluations — score a dataset
Last updated
Was this helpful?