Use this page to choose a detector, input representation, dataset name, and compatible model recipe. The download guide covers published data, while the generation guide covers custom data production.
Choose a representation, then a detector and sample¶
Start with the input representation. Each detector page lists its physics samples, exact TFDS names, schema versions, and compatible model recipes. Pileup conditions distinguish samples within a detector.
| Input representation | Detector or dataset family | Typical use |
|---|---|---|
| Tracks and clusters | CMS, CLD, CLICdet, ColliderML, MAIA, IDEA | Reconstructed tracks and calorimeter clusters; the default starting point for CMS, CLD, and CLIC. |
| Hits | CLD, CLICdet, ColliderML | Tracker and calorimeter measurements for dedicated hit-model research. |
The catalog includes maintained recipes and explicitly identified partial workflows. A builder’s existence does not imply a trained model, physics validation, or public download. Detector pages identify those boundaries. Dataset schema, Python package, and model artifact versions are independent.
After selecting a sample, download and verify published data or follow dataset generation. The shared Key4HEP production guide covers both CLD/CLIC representations.
See detector event views for reusable visualizations of the input representations and reference particles.
Configuration partitions and event splits¶
CMS, CLD, CLIC, MAIA, and ColliderML recipes use configuration partitions 1 through 10; IDEA uses configuration 1. A configuration partition contains its own train and test event splits: the number identifies the partition, while train or test identifies the event split. The standard ten-partition model recipes combine all ten partitions. Configuration 1 provides a one-partition subset for a small experiment; the configured training corpus spans all ten.
The on-disk layout is:
<data root>/<dataset name>/<configuration>/<version>/For example:
data/tfds/tensorflow_datasets/cld/cld_edm_ttbar_pf/1/3.3.0/Pass the detector-level directory containing the dataset-name directories as --data-dir; for this example it is data/tfds/tensorflow_datasets/cld.
Published availability¶
As checked on 28 October 2026, the Hub snapshot includes metadata, feature definitions, and all declared split shards for these current-recipe selections:
| Representation | Detector | Published datasets | Version | Configuration |
|---|---|---|---|---|
| Tracks and clusters | CLD | cld_edm_{ttbar,ww_fullhad,qq,zz}_pf | 3.3.0 | 1 |
| Tracks and clusters | CLICdet | clic_edm_{ttbar,ww_fullhad,qq}_pf | 3.3.0 | 1 |
| Tracks and clusters | ColliderML | colliderml_ttbar_nopu_pf, colliderml_ttbar_pu200_pf | 1.2.0 | 1 |
| Tracks and clusters | IDEA | idea_edm_{ttbar,ww_fullhad,qq}_pf | 0.2.0 | 1 |
| Hits | CLD | cld_edm_{ttbar,ww_fullhad,qq}_hits | 3.2.2 | 1 |
| Hits | CLICdet | clic_edm_{ttbar,ww_fullhad,qq}_hits | 3.2.2 | 1 |
The download guide supplies the CLD 3.3.0 inputs required by
the training guide. Only configuration 1 is published
for the selections above; use --data_config 1 rather than the standard
ten-partition training selection. Produce additional partitions with
the generation workflow.
Treat the live Hub tree as the source of truth for publication availability. The tfds_mapping, tfds_hit_mapping, and model dataset entries in particleflow_spec.yaml define what the current code expects. A usable dataset has a detector, name, configuration, and version that match in both contexts.
For scientific provenance and citations, use the publication map and the dataset record associated with the version you consume.