Supported detectors

Each reader is a registered dataset class; a config selects one with type under data.train, data.val and data.test.

Dataset type	Detector	Source
PILArNetH5Dataset	LArTPC, PILArNet-M simulation	HDF5 shards
PILArNetParquetDataset, PILArNetParquetIterableDataset	LArTPC, PILArNet-M	Parquet, local or streamed from Hugging Face
JAXTPCDataset	LArTPC, JAXTPC simulation	HDF5: 3D deposits, 2D wire signals, correspondence, labels
LUCiDDataset	water Cherenkov, LUCiD simulation	HDF5: PMT response and 3D track segments
LUCiDEventSSLDataset	water Cherenkov	LUCiD events through the pimm-data readers
DefaultDataset	any	one directory of .npy files per event
ConcatDataset	—	several datasets as one

PILArNet-M

Download it with the bundled script:

uv run python scripts/pilarnet/download.py --revision v3                # all splits, about 168 GB
uv run python scripts/pilarnet/download.py --revision v3 --split test   # test split, about 7 GB

Files go to ~/.cache/pimm/pilarnet/v3 unless you pass --output-dir, and the script offers to add PILARNET_DATA_ROOT_V3 to your shell configuration. The reader looks for data in the config’s data_root, then in $PILARNET_DATA_ROOT_<REVISION>, then in ~/.cache/pimm/pilarnet/<revision>.

The reader accepts revisions v3 (the default) and v3_extra, which adds momentum vectors, particle lineage and interaction-vertex tables. It reads v2 as v3, with a warning, and rejects v1. DeepLearnPhysics/PILArNet-M-mini is a 120-event sample in the same layout.

Option	Default	Effect
split	"train"	split directory, or a list of them
min_points	1024	skip events with fewer points
max_len	-1	cap the number of events
energy_threshold	0.13	drop points at or below this energy
remove_low_energy_scatters	False	drop the low-energy-deposit cluster
overlay_n_events	1	overlay several events into one point cloud

JAXTPC

JAXTPCDataset reads co-indexed HDF5 files from a JAXTPC production: seg (3D deposits), resp (2D wire signals), corr (3D-to-2D correspondence) and labl (track-to-label tables). The modalities you load decide what coord is:

Task	modalities	Output
3D segmentation	("seg", "labl")	coord (N, 3), segment (N,)
2D segmentation	("resp", "corr", "labl")	coord (E, 2) from the correspondence, with labels; resp_coord also present
2D self-supervised	("resp",)	coord (M, 2), all wire planes merged, with plane_id
Everything	("seg", "resp", "corr", "labl")	3D coord plus resp_coord and corr_coord

resp and labl without corr give no labels: wire signals map to tracks only through the correspondence. Other options are label_key ("particle", "cluster" or "interaction"), volume and min_deposits. The committed recipe detector/semseg/semseg-pt-v3m2-jaxtpc-5cls reads its root from JAXTPC_DATA_ROOT.

With 2D coordinates, use GridSample, ToTensor, Copy, Collect, RandomDropout, ShufflePoint, RandomJitter, RandomScale, RandomFlip and PositiveShift. RandomRotate, NormalizeCoord and SphereCrop assume three dimensions.

LUCiD

LUCiDDataset reads water Cherenkov simulation output from a root with sensor/ and seg/ subdirectories, named {dataset_name}_sensor_NNNN.h5 and {dataset_name}_seg_NNNN.h5 (or sensor_events_NNNN.h5 and segment_events_NNNN.h5).

Task	modalities	output_mode	Output
Event classification	("sensor",)	"response"	one row per PMT: coord, energy (photoelectrons), time
Per-sensor instance separation	("sensor",)	"labels"	per-particle entries: coord, energy, segment, instance
3D track reconstruction	("seg",)	any	coord (N, 3), energy, time, track_ids, pdg, parent_ids
Joint 3D and sensor	("seg", "sensor")	"separate"	the readers’ raw keys

PMT coord is a 3D position when the sensor reader has PMT positions (pmt_positions, pmt_positions_file, or the file’s config/pmt_positions); otherwise it is a sensor index of shape (N, 1).

data = dict(train=dict(
    type="LUCiDDataset",
    data_root="/path/to/dataset_wc",
    dataset_name="wc",
    modalities=("sensor",),
    output_mode="response",
    transform=[...],
))

Your own format

Bring your own data covers DefaultDataset and writing a reader.