Clone a dataset
Goal. Get a populated database to work against in one command. About two minutes.
Find one
strata hub list-datasets
NAME BADGE LICENSE PRIMITIVES SIZE DOWNLOADS
titanic - CC0-1.0 branches json kv 1.1 MB 231
ab-test-results - CC0-1.0 branches json kv 52.6 kB 8
world-bank-indicators - CC-BY-4.0 json kv 533.8 kB 6
The PRIMITIVES column says which models a dataset uses, and branches there
means it ships with more than one.
Clone it
strata clone titanic ./titanic
cloned titanic@main into ./titanic (11 objects, 1.1 MB)
That is a complete local database. No server, no account, nothing running.
Look at what arrived
strata ./titanic describe
StrataDB 1.2.2 · durable_local
branch default · branches: cleaned, default, raw, test, train · spaces: default
capabilities: arrow event graph_core inference json kv vector vector_index
kv 9 · json 1309 · events 0
Five branches. The dataset ships with raw, cleaned, train and test
as branches of one database, not as four files you have to keep in step.
strata ./titanic --branch raw --raw json count
strata ./titanic --branch cleaned --raw json count
That is the pattern worth noticing: a train/test split is a fork, so the two share storage and their common ancestry is a fact the database knows rather than a naming convention.
Work on it without spoiling it
strata ./titanic branch fork cleaned my-experiment
strata ./titanic --branch my-experiment json set passenger:1 '$.fare' 100
The dataset’s own branches stay as they were shipped. If you want to get back to exactly what you downloaded, delete your branch.
Where it came from
strata ./titanic remote
dataset titanic
branch main
A clone remembers its origin, so a database you find later can say what it is.
What a clone is not
It is a copy, not a subscription. Nothing syncs, nothing updates, and writing to it affects only your copy. Re-clone to get a newer version.
Failure modes
The clone fails or hangs. The hub is a network service; check strata hub list-datasets first to see whether it is reachable at all.
The target directory exists. Clone into a new path.
A dataset’s licence matters. The LICENSE column is there for a reason, and
CC-BY-4.0 has attribution requirements that CC0-1.0 does not.
Next
Combine data models, or Give each run a branch to experiment against what you just cloned.