Datasets

Versioned dataset assets, their file manifests, and the data quality report.

Dataset versions with file manifests and quality metrics.

What you see

Region	What it shows
Versions	Every immutable version, with its checksum
File manifest	What is in the version, with per-file sizes
Preview	Sample records, for formats the console can render — and only where policy permits
Data quality	Computed at push time over the whole dataset
Usage	Which jobs referenced this dataset, which is what makes a version safe to delete

The quality report

Computed once, at push time, over the full dataset — sampled only for very large ones. It counts what a machine can count and does not pretend to judge content.

Metric	Why it is there
Records	The number you thought you uploaded, confirmed
Duplicates	Near-duplicates and byte-identical ones, counted separately. A high rate means the model will overfit to repeated samples
Empty	Records with no usable content
Length p50 / p95	Size max_seq_len from p95, not from the mean — the mean will truncate a fifth of your data
Field drift	Records whose field set differs from the majority, and which fields they are missing

Field drift is the one that catches real bugs. A few hundred records missing the field your loss reads will not fail the job; they will quietly train on nothing and drag the curve.

Protected datasets

A dataset can be governed so that a trainer may use it without being able to read it. On such a dataset the console shows identity, version, schema, lineage and authorisation state — and no preview, no sample records, and no download.

That is enforced on the server, not by hiding buttons. Logs, validation samples and artifacts are covered by the same policy, so they cannot become an alternative export path.

What you can do

• Reference a version in a config or with --train-dataset.
• Push a new version with sf dataset push.
• Check quality before pushing with sf dataset check and sf dataset quality.

See datasets for the full workflow.