mlflow-at1 0.1.0

MLflow artifact repository backed by verified, compressed, byte-exact AT-1 containers

mlflow-at1 — verified, compressed MLflow artifacts

Store MLflow models, checkpoints, and datasets as verified, byte-exact AT-1 containers instead of raw files. Every artifact is compressed losslessly (never-worse vs xz), carries an embedded SHA-256, and is byte-exact-verified on download — a tampered or corrupted artifact is refused, not silently returned.

That makes an artifact in your MLflow store provably the exact bytes you logged — the reproducibility / provenance hook that regulated AI (EU AI Act, NIST AI RMF, FDA 21 CFR Part 11) needs: "this model is the exact model that was trained, untampered."

It's a pure-glue plugin — no engine code ships here. It registers the at1:// artifact-repository scheme and shells out to the at1 CLI.

Install

npm i -g @tinyfiles/cli          # the at1 binary (the codec; no Python engine needed)
pip install mlflow-at1           # this plugin (registers the at1:// scheme)
#   pip install mlflow-at1[full] # if you want full MLflow rather than mlflow-skinny

Use

Point an experiment's artifact location at an at1:// path; everything else is normal MLflow:

import mlflow
exp = mlflow.create_experiment("regulated-model", artifact_location="at1:///data/mlartifacts")
with mlflow.start_run(experiment_id=exp):
    mlflow.log_artifact("model.bin")      # -> stored as a verified, compressed AT-1 container
    # mlflow.sklearn.log_model(...) etc. all route through AT-1 too

# later — download is a byte-exact, SHA-256-verified reconstruction:
path = mlflow.artifacts.download_artifacts(run_id=run_id, artifact_path="model.bin")

at1://<path> is a local/mounted filesystem base directory for the containers (S3/GCS backing is a follow-on). Set AT1_BIN to point at a specific at1 binary if it isn't on PATH.

What you get

• Smaller — lossless, never-worse than xz; typically a real reduction on weights/checkpoints/datasets.
• Verified — every container embeds a SHA-256; download refuses anything that doesn't reconstruct byte-for-byte. Explicit proof per artifact: repo.integrity("model.bin") → integrity OK (sha256 …).
• Addressable — one .at1 per artifact; tabular artifacts stay queryable in place via the at1 CLI.
• Zero lock-in — the container is an ordinary file with a tiny open decoder; you can always get the exact original back, with or without this plugin.

Test

pip install mlflow-skinny pytest && python -m pytest tests/ -q   # needs the `at1` CLI on PATH

5 tests: log→download byte-exact, logical listing, directory trees, integrity + tamper-rejection, delete.