Training

Artifacts

Artifacts are content-addressed files attached to a run: checkpoints, binaries, evaluation dumps. run.artifact(path, step=..., name=...) uploads and records the artifact id as a string metric at those coordinates; run.checkpoint, run.image, run.audio, run.rows are typed wrappers. run.fetch(name) returns the newest file uploaded under a name.

Verify a download

payload = bytes((i * 31) % 256 for i in range(256 * 1024))
expected = hashlib.sha256(payload).hexdigest()

with tempfile.TemporaryDirectory() as tmp:
    path = Path(tmp) / "weights.bin"
    path.write_bytes(payload)

    with launchhelm.init("quickstart", name="artifact-demo") as run:
        artifact = run.artifact(path, step=0, name="weights")
        print("artifact:", artifact.id)
        print("sha256 match:", artifact.spec.get("sha256") == expected)

The artifact document's spec.sha256 is the recorded digest — compare it against your local file to prove the upload; the download on the run's Details tab serves the same object.

Reading telemetry back

Tooling can query points directly with launchhelm.Client. The query example under Examples prints one bounded page (up to limit points) of a stream id — copy the id from the run page. Longer ranges need paging; the query does not return an unbounded result.