CERN/colliderml-benchmark-results

ColliderML Benchmark Results

Machine-scored leaderboard results for the ColliderML benchmark tasks.

Structure

results/
  {task}/
    {username}/
      {predictions_sha256}.json

Each JSON file records one scored submission:

Field	Description
submission_id	UUID assigned by the backend
task	Benchmark task name (e.g. tracking)
submitter	HuggingFace username
model_repo_id	Optional link to the submitter's model repo
submitted_at	ISO 8601 timestamp
scores	Dict of metric name → value
credits_earned	Credits awarded for this submission
is_baseline	Whether this is a reference baseline
predictions_sha256	SHA-256 of the submitted parquet

Tasks

Task	Primary metric	Dataset
tracking	TrackML weighted efficiency	ttbar_pu200
jets	Jet classification AUC	ttbar_pu200
anomaly	Anomaly AUROC	mixed
tracking_latency	Inference latency (ms)	ttbar_pu200
tracking_small	TrackML eff under parameter budget	ttbar_pu200
data_loading	Load throughput (events/s)	ttbar_pu0

Integration

Results are pushed automatically by the ColliderML backend after each scored submission. The ColliderML leaderboard Space reads from this dataset to render the public table.

Models that include .eval_results/colliderml_*.yaml files link back to this dataset, so scores appear on model cards automatically.

Links

• Library: OpenDataDetector/ColliderML
• Data: CERN/ColliderML-Release-1
• Docs: opendatadetector.github.io/ColliderML