Panda#

Panda is a sparse hierarchical point encoder for LArTPC events, pretrained on unlabeled data with a multi-view, prototype-based self-distillation objective. Its backbone is PT-v3m2, a Point Transformer V3 variant; its pretraining model is Sonata-v1m1, pimm’s implementation of Sonata. Task heads on the encoder label points, find particles and group interactions. Paper: arXiv:2512.01324.

Released checkpoints#

Checkpoint	Model	Output
`Panda-Base`	`PT-v3m2` encoder	a `Point`; `point.feat` holds per-point features
`Panda-Semantic`	`DefaultSegmentorV2` on `PT-v3m2`	`seg_logits`, shape (N, 5)
`Panda-Particle`	Panda detector on `PT-v3m2`	query masks and classes; `model.postprocess()` turns them into per-point particles
`Panda-Interaction`	Panda detector on `PT-v3m2`	query masks; `model.postprocess()` turns them into per-point interactions

Semantic classes, in output order: shower, track, Michel, delta, low-energy deposit.

Particle classes, in output order: photon, electron, muon, pion, proton. Points that belong to no particle, such as low-energy deposits, get instance -1.

Input pipeline#

The Panda recipes apply these transforms at inference:

from pimm.datasets.transform import Compose

panda_transform = Compose([
    dict(type="NormalizeCoord", center=[384.0, 384.0, 384.0], scale=768.0 * 3**0.5 / 2),
    dict(type="LogTransform", min_val=0.01, max_val=20.0, keys=("energy",)),
    dict(type="GridSample", grid_size=0.001, hash_type="fnv", mode="train", return_grid_coord=True),
    dict(type="ToTensor"),
    dict(type="Collect", keys=("coord", "grid_coord"), feat_keys=("coord", "energy")),
])

NormalizeCoord maps the PILArNet-M volume, a cube of side 768 centered at 384, into the unit ball.
LogTransform maps energy onto [-1, 1] on a log scale between 0.01 and 20.
GridSample keeps one point per occupied cell of a 0.001 grid and adds integer grid_coord. With mode="train" it picks a random point in each cell; seed NumPy to make the choice repeatable. Predictions refer to the points it keeps.
Collect builds feat from coordinates then energy, so the backbone has in_channels=4.

Recipes#

Recipe	Trained
`panda/pretrain/pretrain-sonata-v1m1-pilarnet-smallmask`	the Panda encoder, from scratch, with self-distillation
`panda/pretrain/pretrain-sonata-v1m1-pilarnet-smallmask-v3m8`	the same with the `PT-v3m8` backbone
`panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-lin`	semantic head on a frozen encoder
`panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-dec`	decoder and head on a frozen encoder
`panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-fft`	the whole semantic model
`panda/semseg/semseg-pt-v3m2-pilarnet-ft-5cls-scratch`	the whole semantic model; run it without weights
`panda/panseg/detector-v5-pt-v3m2-ft-pid-{dec,fft,scratch}`	the particle detector
`panda/panseg/detector-v5-pt-v3m2-ft-vtx-{dec,fft,scratch}`	the interaction detector
`panda/panseg/detector-v5-pt-v3m2-ft-{pid,vtx}-fft-detector`	the whole detector, starting from the released Particle or Interaction weights pinned in the recipe

To fine-tune the semantic and detector recipes from Panda-Base, pass --train.weight hf://DeepLearnPhysics/Panda-Base; see Fine-tune. The detector recipes build detector-v5. The repository also registers detector-v3m2, detector-v4 and detector-v5m2.

Cite#

@misc{young2025pandaselfdistillationreusablesensorlevel,
  title         = {Panda: Self-distillation of Reusable Sensor-level Representations for High Energy Physics},
  author        = {Samuel Young and Kazuhiro Terao},
  year          = {2025},
  eprint        = {2512.01324},
  archivePrefix = {arXiv},
  primaryClass  = {hep-ex},
  url           = {https://arxiv.org/abs/2512.01324}
}

Panda builds on Sonata and Point Transformer V3; their entries are on Citing pimm.