juniper-data

A FastAPI service that generates, versions, and serves ML datasets as NPZ artifacts.

juniper-data turns a catalogue of dataset generators into a REST service: you ask it for a dataset by name and parameters, and it returns a versioned, NPZ-formatted train/test/full split. The catalogue spans synthetic classification problems (two-spiral, concentric circles, XOR, Gaussian mixtures, moons, checkerboard), image sets (MNIST), the ARC-AGI visual-reasoning families, a CSV/JSON import path, and a family of time-series and irregularly-sampled sequence generators (autoregressive, Mackey-Glass, multi-sine, delay-product, equities, and the irregular-Δt equities_seq contract). A named-version registry, tag filtering, batch creation, and per-dataset preview round out the surface. Call GET /v1/generators for the live catalogue.

It is the foundational data layer of the platform: the dataset identifiers it returns are the substrate juniper-cascor trains on and juniper-canopy visualises.

Part of the Juniper platform. juniper-data is the dataset-generation service of Juniper — a multi-package ML research platform built around constructive (Cascade-Correlation) and recurrent neural networks. It runs standalone; the rest of the platform consumes it over HTTP (see juniper-data-client).

Install

pip install juniper-data            # from PyPI

For development from a clone (the optional extras are api, arc-agi, equities, mnist, observability, test, dev, all):

git clone https://github.com/pcalnon/juniper-data.git && cd juniper-data
pip install -e ".[all]"

MNIST / Fashion-MNIST (optional extra)

The mnist generator loads the real MNIST / Fashion-MNIST datasets from the Hugging Face Hub and needs the (heavy) datasets chain, shipped behind an explicit extra — it is never part of the base install:

pip install "juniper-data[mnist]"

• First call downloads from the Hub into the Hugging Face cache (HF_HOME, default ~/.cache/huggingface; the Docker image pins it to /app/data/hf-cache so a mounted data volume persists it). Later calls are served from the cache.
• Offline deployments must seed that cache ahead of time (run one generation for each dataset while online, or copy a populated HF_HOME in); with HF_HUB_OFFLINE=1 the generator then works entirely from the cache.
• Without the extra installed, the generator is unavailable: the registry reports available: false and POST /v1/datasets returns 501 with the install hint instead of a masked 500.
• The service Docker image ships the extra (it is compiled into requirements.lock), so MNIST generation works in containers out of the box.

Run

uvicorn --factory juniper_data.api.app:get_app --reload    # binds 127.0.0.1:8100
curl http://localhost:8100/v1/health/ready
curl http://localhost:8100/v1/generators                   # the live generator catalogue

Create a dataset over the REST API:

curl -sX POST localhost:8100/v1/datasets \
  -H 'Content-Type: application/json' \
  -d '{"generator": "spiral", "name": "demo", "params": {"n_spirals": 2, "noise": 0.1}}'

Or generate one in-process, without the service:

from juniper_data.generators import SpiralGenerator, SpiralParams

dataset = SpiralGenerator.generate(SpiralParams(n_spirals=2, n_points_per_spiral=100, noise=0.1))
# dataset: dict of float32 arrays — X_train, y_train, X_val, y_val, X_test, y_test

Data contract

Datasets are NPZ archives with the keys X_train, y_train, X_val, y_val, X_test, y_test, all float32. This is the contract every Juniper consumer reads.

Configuration

Settings load from the JUNIPER_DATA_ environment namespace (juniper_data/api/settings.py) and honor the Docker _FILE secret convention. The most common knobs (full surface in docs/REFERENCE.md):

Docker

docker build -t juniper-data:latest .
docker run --rm -p 8100:8100 -e JUNIPER_DATA_HOST=0.0.0.0 juniper-data:latest

Multi-stage build (Python 3.14-slim); health is probed at /v1/health/ready. For the full stack, see juniper-deploy.

Status

Live on PyPI. The current version is shown by the badge above; see CHANGELOG.md. Consumed by juniper-cascor and juniper-canopy via JUNIPER_DATA_URL, and by juniper-data-client programmatically.

Documentation

• docs/QUICK_START.md — get running in five minutes
• docs/USER_MANUAL.md — comprehensive usage guide
• docs/api/JUNIPER_DATA_API.md — full REST reference (filtering, batch, tagging, versioning)
• docs/REFERENCE.md — configuration and environment-variable reference
• docs/DOCUMENTATION_OVERVIEW.md — index of all juniper-data docs

License

MIT — see LICENSE.