colliderml 0.4.0rc1

A modern machine learning library for high-energy physics data analysis

ColliderML

A modern machine learning library for high-energy physics data analysis.

Installation

pip install colliderml                 # core + Polars loader + unified load()
pip install 'colliderml[sim]'          # local simulation (needs Docker/Podman)
pip install 'colliderml[remote]'       # SaaS backend client
pip install 'colliderml[tasks]'        # benchmark task reference baselines
pip install 'colliderml[all]'          # everything above + dev tools

For development: pip install -e ".[dev]"

Getting the data

Option 1 — Python one-liner (downloads on first call, then caches):

import colliderml

frames = colliderml.load("ttbar_pu0", max_events=200)
print(frames["particles"])             # Polars DataFrame

Option 2 — CLI (explicit download, then load with the library):

colliderml download --channels ttbar --pileup pu0 --objects particles,tracker_hits,calo_hits,tracks --max-events 200

Cache location: default ~/.cache/colliderml, or set COLLIDERML_DATA_DIR. List downloaded configs: colliderml list-configs.

Option 3 — HuggingFace only:

from datasets import load_dataset
dataset = load_dataset("CERN/ColliderML-Release-1", "ttbar_pu0_particles", split="train")

Running simulations

New in v0.4.0: generate events yourself with the full ODD pipeline, either locally in a container or via the SaaS backend.

import colliderml

# Local: runs inside the OpenDataDetector software container.
# Needs Docker or Podman; the `[sim]` extra provides the driver.
result = colliderml.simulate(preset="ttbar-quick")
print(result.run_dir)                  # parquet outputs land here

# Remote: submit to the SaaS backend (requires an HF token).
# The `[remote]` extra pulls in requests; no container runtime needed.
result = colliderml.simulate(preset="higgs-portal-quick", remote=True)
print(result.remote_request_id)

CLI equivalents:

colliderml list-presets
colliderml simulate --preset ttbar-quick --local
colliderml simulate --preset higgs-portal-quick --remote
colliderml status <request-id>
colliderml balance

See the Local Simulation and Remote Simulation guides for details.

Benchmark tasks

New in v0.4.0: six built-in benchmark tasks — tracking, jets, anomaly, tracking_latency, tracking_small, and data_loading — with a unified registry and a leaderboard backed by the SaaS backend.

import colliderml.tasks

print(colliderml.tasks.list_tasks())
scores = colliderml.tasks.evaluate("tracking", "my_preds.parquet")
colliderml.tasks.submit("tracking", "my_preds.parquet")   # earn credits on new bests

Reference baselines (scikit-learn for BDT/IsoForest) ship with the [tasks] extra. See the Benchmark Tasks guide for details.

Using the library

The notebook notebooks/colliderml_loader_exploration.ipynb shows the data-loading and analysis helpers: load_tables, exploding event tables, pileup subsampling, calibration, and plotting.

Full docs: https://opendatadetector.github.io/ColliderML

Development

pytest -v -m "not integration"

Docs are built with VitePress: npm ci --prefix docs && npm run --prefix docs docs:build.

License

MIT License