Documentation

Quickstart

Search is free and needs no account — explore the catalog and open any dataset for its full record: origin, licensing, structure, and lineage as documented at the source.

from archivum import Client

client = Client()
ds = client.pull("clinical-notes-mimic-derived", version="v3.2.0")

print(ds.coverage.total)     # % documented
print(ds.license.spdx)       # licence as published
ds.export("llamaindex")      # provenance travels with the data

Every record links back to its source platform. Archivum stores metadata only — downloads always happen at the origin, under the origin’s terms.

Coverage methodology v1.0

Documentation Coverage measures one thing: how much of a dataset’s provenance was documented at the source when Archivum checked. It is a factual measure of the record, not a judgment of the dataset. Twenty-eight checks are grouped into four equally weighted sections; each check asks a question with a verifiable yes-or-no answer.

How each check is scored

• Documented — 1 point. Archivum retrieved the artifact itself through the platform’s API: a licence field, a file manifest, a schema, a commit history.

• Reported — half a point. The publisher stated it in prose — a README section, a dataset card paragraph — that Archivum retrieved but did not independently confirm.

• Not found — 0 points. Absent from the published metadata at the time of the check. This is a fact about the record, not a defect in the data. Checks that cannot apply on a platform are excluded from that dataset’s denominator entirely.

A section’s score is its points divided by its applicable checks. The overall figure is the mean of the four sections. That is the entire calculation — no weights to tune, no popularity bonus, no judgment call.

The 28 checks

Coverage reflects what was present at the source on the date shown with each record. It describes documentation, not data quality, and nothing on this site is legal advice — read the licence at the origin before commercial use.

Lineage model

Lineage answers one question: what does the source document about this data’s path from origin to the current version? Archivum models the chain as stages — original source, raw acquisition, cleaning, annotation, and current version. Each stage the source documents records the actor, a content fingerprint where available, a timestamp, and whether it was documented or reported.

Stages the source does not document render as explicit gaps — dashed, in warning color — on every lineage graph. Archivum never invents a stage to close a gap. The gap is the information.

API reference

The read API mirrors what the site shows — the same records, the same evidence labels. All endpoints are read-only and return JSON.

Responses carry the coverage rules version so pipelines can pin against it.