DataFolio

A small, self-documenting home for the data behind one analysis.

You already have the objects: a feature table, some labels, an array of embeddings, a fitted model, a parameter dict, a QC plot. DataFolio gives them one directory, one name each, and one sentence each explaining what they are — then hands them back by name months later, from any notebook.

from datafolio import DataFolio

folio = DataFolio("analysis/experiment-12")

folio.add("features", features, description="One row per neuron; normalized morphology")
folio.add("labels", labels, description="Manual labels after the March review")
folio.add("params", {"alpha": 0.1, "seed": 7}, description="Settings used for the paper figures")
folio.add_model("classifier", clf, inputs=["features", "labels"],
                description="Cell-type classifier on reviewed labels")

# A different notebook, six months later. One path, no filenames.
folio = DataFolio("analysis/experiment-12")
folio.describe()

DataFolio: analysis/experiment-12
=================================

Created: March 14, 2026 at 2:05 PM EDT
Updated: March 20, 2026 at 11:42 AM EDT

Tables (2):
  • features: One row per neuron; normalized morphology
    ↳ size: 70.2 KB
  • labels: Manual labels after the March review
    ↳ size: 8.3 KB

JSON Data (1):
  • params: Settings used for the paper figures
    ↳ type: dict
    ↳ size: 31 B

Models (1):
  • classifier: Cell-type classifier on reviewed labels
    ↳ size: 1.4 KB
    ↳ inputs: features, labels

Every item, what it is, and what it was made from — without opening a file. The same call works on a folio someone sends you as a cloud path.

features = folio.get("features")
clf = folio.get_model("classifier")

Install

pip install datafolio            # core
pip install 'datafolio[polars]'  # + lazy scans and Polars frames

Python 3.10+.

Who it's for

One researcher, or a small team sharing mostly read-only work, holding a few to a few dozen understandable objects per analysis. It assumes you are comfortable in pandas, Polars, numpy, and scikit-learn and do not want another framework between you and them.

Good fit when you want to stop writing pd.read_parquet(BASE / "features_v3_reviewed.parquet") in every notebook, remember which of five similar files the paper used, keep one analysis's objects together, or hand a colleague a single path.

What you get

One write path, one read path. add() picks the storage format from the object's type; get() reads the catalog and returns the matching Python object.

folio.add("table", dataframe)            # pandas / Polars / LazyFrame -> Parquet
folio.add("counts", series)              # pandas / Polars Series -> Parquet
folio.add("embeddings", array)           # numpy -> .npy
folio.add("params", {"alpha": 0.1})      # dict / list / tuple / set / scalar -> JSON
folio.add_model("classifier", clf)       # -> joblib (or skops)
folio.add_file("plots/qc.png")           # -> the file, unchanged
folio.reference_table("raw", "gs://lab-data/raw.parquet")   # link, don't copy

Descriptions live next to the data. The most valuable metadata is usually one sentence saying which of several near-identical files this is. It is stored in the catalog, shown by describe(), and readable by people who never install the package.

Ordinary files, not a container format.

experiment-12/
├── items.json                    # the authoritative catalog (+ metadata + snapshots)
├── CONTENTS.md                   # derived, human-readable inventory
├── README.md                     # how to read this directory without datafolio
├── tables/features--r2.parquet
├── models/classifier--r8.joblib
└── artifacts/params--r5.json

Local or cloud, same code. DataFolio("gs://team-analysis/experiment-12").

Lineage, snapshots, and a CLI when you need them:

folio.add("predictions", preds, inputs=["classifier", "holdout"])
folio.create_snapshot("paper-v1", description="Figures 2–4 in the submission")

datafolio describe
datafolio snapshot status

What it is not

Not a database or query engine, not a workflow orchestrator, not data version control, not an experiment tracker, not a multi-writer system, not a backup. Many readers, one writer — treat a cloud folio as single-writer. Comfortable at a few dozen items per folio, not thousands.

DataFolio's rule is to be as lightweight as possible and hand off to better tools as soon as possible: big tables go to Polars as a lazy scan, or to you as a path. See What DataFolio is not for the full list, including the sharp edges worth knowing before you rely on it.

Documentation

Read in this order:

1. Your first folio — ten minutes
2. Everyday patterns
3. Tables: big, external, and lazy
4. Models
5. Sharing a folio
6. Snapshots
7. Reading a folio without DataFolio
8. What DataFolio is not

Reference: API cheat sheet · CLI · Migrating from 1.x

Development

uv sync
poe test                      # pytest with coverage
uv run ruff check src/ tests/
poe doc-preview               # local docs server

License

MIT