CERN/anomaly_detection_cmsl1t

Trigger Anomaly Detection for New Physics at the Large Hadron Collider

This dataset is a mirror of the Zenodo record: https://doi.org/10.5281/zenodo.21787779

This dataset contains Level-1 Trigger objects from the CMS experiment at the CERN Large Hadron Collider, assembled for research on unsupervised anomaly detection in the trigger. The goal of unsupervised anomaly detection in this context is the discovery of new physics. This data set does not contain new physics. It is meant for research on anomaly detectors.

In the new physics search context, recording true anomalies or simulating them is impossible, compared to other settings (e.g. industrial applications of anomaly detection) where this is commonly done. Anomaly simulation data sets are provided, but they should be used with the aforementioned caveat in mind. Aside from its high statistics, this data set uniquely provides simulations of the normal data.

• normal data: zero-bias events recorded during 2025 proton–proton running (runs 396102, 398183); these events are chosen at random from all the proton-proton collisions that happen inside the CMS detector and are recorded by it.

• anomaly simulations: 20 simulated signal data sets covering Higgs, multi-Higgs, SUSY and exotic scenarios from the CMS Run 3 Winter25 campaign.

• normal data simulation: one simulated zero-bias-like background sample (SingleNeutrino).

The normal data has 20,887,636 events. The normal data simulation has 2,000,000 events. The 20 anomaly simulations amount to 13,012,931 events.

Each event provides particle level and event information, as recorded by the trigger. The data is published pre-partitioned into training, validation and test splits: the zero-bias data 60/20/20, the simulated samples 60/40 between validation and test. All the feature values are in the trigger-native format of hardware integers.

Specific technical data can be found in the github repo that was used to produce this data: https://github.com/bb511/adl1t_datamaker The data were produced at commit 16dcaac of that repository.

Comparison with other trigger algorithms

The level-1 trigger menu contains hundreds of algorithms that take data, like the one present in this record, and output a decision. All of them ran on every event of the presented data sets, outputting a decision. The results are stored in the seeds folder and can be used to do comparative studies between your anomaly detection algorithm and the standard algorithms running in the CMS trigger. The trigger's own anomaly detection algorithms, L1_AXO_* and L1_CICADA_*, are left out, since benchmarking an anomaly detector against the decisions of another anomaly detector would be circular. The folder also contains an L1bit field, which encodes the logical OR of the algorithm columns deposited beside it. Skip the seeds folder if you want kinematics alone and do not want to compare your anomaly detector with the rest of the algorithms.

Row i of <data set>-seeds is row i of <data set>, which is the correspondence that always holds. Neither column is a key on its own: every event the standard preprocessing dropped carries order = -1, and event cycles over a hundred values in ggH-suep-decay and smj-case-A. Among the rows with order >= 0 it is unique within a split, so a filter or a shuffle can be undone by joining on order once those rows are set aside.

Loading

Every data set is a configuration of its own, and every one of them has a -seeds twin holding the other algorithm decisions for the same events in the same order.

from datasets import load_dataset
normal = load_dataset("podagiu/anomaly_detection_cmsl1t", "ZB_run396102", split="train")
signal = load_dataset("podagiu/anomaly_detection_cmsl1t", "WtoTauto3Mu", split="validation")
menu = load_dataset("podagiu/anomaly_detection_cmsl1t", "WtoTauto3Mu-seeds", split="validation")

HuggingFace calls the middle split validation while our files call it 'valid'. We also include a dataloader that we used for performing anomaly detection trigger development on this data. It reads the data, applies our cuts and normalisation, and stacks the collections into pytorch tensors that can be directly fed into models.

import sys
from huggingface_hub import snapshot_download
from hydra import compose, initialize_config_dir
from hydra.utils import instantiate
record = snapshot_download("podagiu/anomaly_detection_cmsl1t", repo_type="dataset")
sys.path.insert(0, record)  # the configs name loader.*, so the record has to be importable
with initialize_config_dir(config_dir=record + "/configs", version_base=None):
    cfg = compose("config", overrides=["paths.root_dir=" + record])
data = instantiate(cfg.data)
data.prepare()
train = data.load("train")

prepare reads the whole record and caches each stage under ./cache, which the ADL1T_CACHE environment variable moves elsewhere; allow it a few times the record's size on disk. train then has x, the model input, of shape (events, 39, 3) under the basis configuration, its padding mask, whether the event passed ANY other algorithm in the trigger l1bit and the label y (0 for zerobias, > 0 for anomalies, < 0 for zerobias simulation). data.load_aux("valid") returns the same for every simulated sample. The pipeline needs python 3.10 or newer with awkward, pyarrow, numpy, torch, omegaconf and hydra-core, which pip install -r requirements.txt at the record's root installs. If you would just rather read the raw data, you need none of the above dependencies. Just call load_dataset.

Layout

data/<data set>/<split>-NNNNN-of-NNNNN.parquet
data/<data set>/seeds/<split>-NNNNN-of-NNNNN.parquet
loader/
configs/
requirements.txt

One row is one event. The four object collections are jagged, holding one entry per in-time object up to the global trigger's capacity of 8 muons, 12 jets, 12 e-gammas and 12 taus, with no padding and no truncation. The energy sums are collections too, of one entry each. The event information, the trigger's verdict and every seed are plain values.

Columns

column	holds
<collection>_<branch>	one collection's features, e.g. muons_muonIEt or jets_jetIEta. The collections are muons, jets, egammas, taus and the energy sums ET, HT, MET, MHT, FET, FHT, and the branch names are the trigger's own
run, lumi, event, bx, orbit, time, nPV_True	event information, carried without a prefix
split, order	the published partition and the position within it, described below
L1bit	whether the trigger accepted the event, i.e. the OR over the algorithm columns deposited in the -seeds config
dataset	the data set the row came from, so that a concatenation stays self-describing
label	0 for zero bias, negative for a simulated background, positive for a signal

A column prefixed by a collection is a list, and every other column is a plain value. The collections have one entry per object in the event. The energy sums have a single entry (list with one value). Everything else contains one value: row["L1bit"] is True, row["event"] is an integer, and ds.filter(lambda r: r["L1bit"]) selects the events the trigger accepted. A -seeds config has one boolean column per trigger algorithm and four other columns: L1bit, dataset, event and order. The trigger's own anomaly detection algorithms, L1_AXO_* and L1_CICADA_*, are left out, since benchmarking an anomaly detector against the decisions of another anomaly detector would be circular.

Data sets

config	label	train	valid	test
GluGluHTo2B_Par-MH-125	1		293,231	195,736
GluGluHto2G_Par-MH-125	2		598,666	399,132
GluGluHto2G_Par-MH-90	3		443,336	295,444
GluGluHto2Tau_Par-MH-125	4		35,706	23,805
GluGlutoHHto2B2WtoLNu2Q_Par-c2-0-kl-1-kt-1	5		179,378	119,714
HHHto4B2Tau_Par-c3-0-d4-0	6		1,200,063	800,123
HHHto6B_Par-c3-0-d4-0	7		383,502	255,858
SUSYGluGluToBBHTo2B_Par-M-1200	8		29,588	19,676
SUSYGluGluToBBHToBB_Par-M-120	9		299,996	200,004
SUSYGluGluToBBHToBB_Par-M-350	10		119,971	80,029
SUSYGluGluToBBHToBB_Par-M-600	11		30,000	20,000
SingleNeutrino_E-10-gun	-1		1,200,002	799,998
TTHTo2C_Par-MH-125	12		1,199,636	800,464
TTHto2B_Par-MH-125	13		1,200,165	800,225
VBFHTo2C_Par-MH-125	14		299,566	199,720
VBFHto2B_Par-MH-125	15		299,564	199,733
VBFHto2Tau_Par-MH-125	16		955,535	636,965
WtoTauto3Mu	17		60,002	39,998
ZB_run396102	0	6,220,470	2,071,292	2,072,007
ZB_run398183	0	6,312,116	2,106,228	2,105,523
ggH-suep-decay	18		59,640	39,760
haa-4b-ma15	19		59,942	39,958
smj-case-A	20		59,458	39,642

Units

Each feature of each object has values that are integer hardware units, as the trigger produces them. Nothing in the files is scaled. Multiply by the following to get GeV, radians and pseudorapidity:

collection	Et	eta	phi
muons	0.5 GeV	0.010875	0.0109083 rad
jets	0.5 GeV	0.0435	0.0436332 rad
e-gammas	0.5 GeV	0.0435	0.0436332 rad
taus	0.5 GeV	0.0435	0.0436332 rad
ET, HT	0.5 GeV
MET, MHT, FET, FHT	0.5 GeV		0.0436332 rad

The decimals are rounded. The steps are exact fractions: calorimeter eta is 0.0870/2, muon eta is 0.0870/8, calorimeter phi is 2pi/144, and muon phi is 2pi/576. muons_muonIEtaAtVtx and muons_muonIPhiAtVtx take the same scales as muon eta and phi. Three energies are missing from the table: muons_muonIEtUnconstrained is 1 GeV per unit rather than 0.5, ET_ETTEM takes the same 0.5 GeV as ET_Et, and jets_jetRawEt has no documented scale, but it's probably 1 GeV per step. The muon energies also carry an offset, since the hardware 0 marks the absence of a muon: the momentum is (muons_muonIEt - 1) x 0.5 GeV and the unconstrained momentum is (muons_muonIEtUnconstrained - 1) GeV. Quality, charge, isolation, index and tower-count fields are already integers and unscaled, as is every seed and every event information field except nPV_True, which is a float32 average per luminosity section in the zero bias and in haa-4b-ma15, and an integer count in the other simulations.

Caveats

Ordering. Objects arrive in the trigger's readout order, which is ET-descending for the calorimeter objects. This is not the case for the muons, except in haa-4b-ma15, whose overlay sorted every collection. The shipped loader sorts objects by ET before processing them.

haa-4b-ma15 carries overlaid pile-up. The sample was simulated without pile-up, and each of its events is the merge of the simulated event with one zero-bias event of the same split, at the level of the trigger objects. The partner events were drawn without replacement with NumPy's PCG64 generator seeded with 42 from the two zero-bias runs concatenated as ZB_run396102 then ZB_run398183; they remain in the zero-bias splits. All the drawn events come from val and 'test', for each stage. The four collections are concatenated, ordered by Et and cut at the trigger's capacity of 8 muons, 12 jets, 12 e-gammas and 12 taus; ET, HT and the tower count are added and clipped at their all-ones code; MET, MHT, FET and FHT are added as vectors and quantised back to their codes. The sample only approximates a simulation with pile-up. Its seeds are the OR of the two events' decisions over the simulation menu and L1bit the OR of those seeds. Its event_info is the zero-bias partner's: the sample has real beam coordinates and a float32 nPV_True.

The menu differs between data and simulation. Zero bias data has 178 algorithm columns and the simulations have 158, of which 145 are shared. Additionally, the order of the other trigger algorithm decisions is not the same between zerobias and simulations.

nPV_True has two types. It is float32 in zero bias and in haa-4b-ma15, and int32 in the other simulations.

Simulation has no beam coordinates. Every simulated sample except haa-4b-ma15 has run of 1 and bx and orbit of 4294967295, the all-ones code of 32 bits, in place of the values a collision would have. On Zenodo orbit holds the all-ones code of its 64 bits instead, 18446744073709551615, which the Hub's viewer cannot render; this is the only place the two copies differ. The zero bias has non-trivial values in these fields, and haa-4b-ma15 has those of its zero-bias partner.

jetRawEt is zero throughout the zero-bias data. The branch is unfilled in original data ntuples, though it has real values in simulation. In haa-4b-ma15 it is zero for the jets that came from the zero-bias partner.

Standard preprocessing

Multiple studies were done internally at CERN on this data set. A number of conventional preprocessing steps were applied in each of these studies. Therefore the record contains two columns that the raw data does not: split, so a file separated from its directory is still self-describing, and order, the position in that ordering, which is -1 for the events that the conventional preprocessing removed. Links to the papers detailing these studies will be attached here once these studies become public.

The split was drawn once with NumPy's PCG64 generator seeded with 42, over the two zero-bias runs concatenated in the order ZB_run396102 then ZB_run398183.

A split can span two configs. The two zero-bias runs were permuted together. Their training rows interleave and order counts across the whole split rather than within one run. To rebuild the same order as in previous studies, read both zero-bias configs, concatenate them, then stable-sort by order with the -1 rows left at the end. Concatenating one run after the other gives the right rows in the wrong order.

The pipeline in loader/ goes through the standard preprocessing steps. In the four stages the studies used: read the tables into one array per collection, drop the events saturated in ET and remove the saturated objects, fit the normalisation on the training split alone and apply it to every other split, then pad each collection to a fixed number of constituents and stack them into one tensor. That object cut is Et < 511 for muons, e-gammas and taus, Et < 2047 for jets and Et < 4095 for MET or FET, each the all-ones code of the object's own energy width, so this cut removes saturated objects. The events cut by this pipeline have the order set to -1: run over the zero bias it removes the 457 events marked -1, 0.0022% of them.

Provenance

Zero-bias data: CMS, 2025, runs 396102, 398183. Simulated samples: CMS Run 3 Winter25 campaign; haa-4b-ma15 from its no-pile-up production (142XnoPU), with pile-up overlaid from the zero-bias data as the caveats describe. The values are the Level-1 trigger's own reconstructed objects rather than offline reconstruction.

Citation

Cite the Zenodo record that this dataset mirrors. The data descriptor is not published yet; its citation will be added here once it is.

@dataset{cms_l1t_anomaly_2026,
  author    = {{CMS Collaboration}},
  title     = {Trigger Anomaly Detection for New Physics at the Large Hadron Collider},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.21787779},
  url       = {https://doi.org/10.5281/zenodo.21787779}
}

Licence

CC0 1.0, a public domain dedication with no restrictions on reuse. See LICENSE. Citation by DOI is requested as a courtesy, not required.

Contact

Questions and problems are welcome as a discussion on this dataset's page, or as an issue on the repository that produced it: https://github.com/bb511/adl1t_datamaker.

Release approval

The public release of this data set was approved at the CMS Collaboration Board meeting of 15 May 2026: https://indico.cern.ch/event/1683535/