• JHEP https://doi.org/10.1007/JHEP08%282022%29231
• arXiv https://arxiv.org/abs/2205.09876
• Zenodo https://doi.org/10.5281/zenodo.6827724

You do not need to download the complete datasets to use this repository. But you may wish to, for example, to train and test your own models.

This repository includes:

• all data to reproduce all plots in the paper (serialized results, not raw input data),
• serialized versions of our trained BDT and NN models,
• some small example datasets. Larger datasets are shared on Zenodo. Thanks for hosting them, Zenodo!

Each run of MadGraph produces an lhe file which includes partonic truth and various metadata. For each model (meaning each lambdaPV or rotated coupling matrix), we are one example lhe file in the

• lhe files dataset.

We share complete train, validation, and test datasets for

• truth-jet, reco-jet, and rotated truth-jet data.

The calo-image datasets are very large. We share two examples:

• calo-image datasets for lambdaPV=1, and
• calo-image datasets for the standard model.

Use a Linux environment with a recent version of python 3.

For compatibility, it may be helpful to start from a clean conda environment.
The bundled version of MadGraph does not work with python 3.10.
We also require gfortran for MadGraph and texlive for plotting.

conda create -n hunting-vampires -c conda-forge python==3.9.12 gfortran==12.1.0

Then to set up

conda activate hunting-vampires

Some LaTeX context is also required for matplotlib, but I have failed to install it through conda (texlive-core doesn't work). From a CERN-linked environment, for example, you can link texlive with export PATH=/cvmfs/sft.cern.ch/lcg/external/texlive/2020/bin/x86_64-linux:$PATH.

All plots used in the paper (and some others) are produced from serialized data which are included in this repository.
To jump into the plotting environment and reproduce those plots, execute:

source example_plots.sh

Choosing $\lambda_\textrm{PV} = 0.5$ as an example.
Again, this hops into an environment.
It then:

• generates the mSME model for MadGraph,
• modifies the MadGraph code to evaluate matrix elements in the lab frame(*),
• runs MadGraph to simulate 3-partonic-jet events under our kinematic selections (4-jet is very slow), and
• moves the lhe to the local directory as pv_msme_0p5.lhe.gz.

(*) MadGraph is modified by the --lab arguments to lib/madcontrol.py. The modification is implemented in liv/use_lab_frame.py, and implements a Lorentz boost in each liv/process/${PROCESS}/SubProcesses/*/auto_dsig?.f file.

We also leave lots of mess behind in the liv/ directory; you can use git diff to see what's there.

Execute:

source example_pv_msme_lhe.sh 0.5

For later use, we convert parts of the lhe file (XML) to an efficient format.

This script first converts it to padded four-momenta in an h5 file pv_msme_0p5_truth.h5.
It then applies kinematic cuts to produce pv_msme_0p5_truth_cut.h5.
It then converts that to the image representation in pv_msme_0p5_truth_cut_images.h5.

Execute:

source example_truth_jet.sh pv_msme_0p5.lhe.gz

(Note that the 0.5 above has become 0p5 here.)

An example lhe (from the test set) with $\lambda_\textrm{PV} = 1$ and up to four partonic jets is included as example_results/pv_msme_3j_4j_1_seed_80.lhe.gz.
Extract it by executing:

source example_truth_jet.sh example_results/pv_msme_3j_4j_1_seed_80.lhe.gz

Lead a serialized BDT model and test it against the $\lambda_\textrm{PV} = 1$ sample that we (could have) generated above.

source example_bdt.sh

This prints out a json-formatted report of its results, in which:

• ntest is the number of testing data,
• log_r_test is the model-versus-symmetry $\log$-likelihood ratio, which equals $nQ$, and
• quality is $Q$ with its standard mean and standard deviation estiamtes.

Many other models are saved in results/models/.
Modify the paths given in example_bdt.sh as arguments to example_bdt.py to test them, too!

Lead a serialized NN model and test it against the $\lambda_\textrm{PV} = 1$ sample that we (could have) generated above.

source example_nn.sh

Just as for the BDT, this prints out a json-formatted report of its results, in which:

• ntest is the number of testing data,
• log_r_test is the model-versus-symmetry $\log$-likelihood ratio, which equals $nQ$, and
• quality is $Q$ with its standard mean and standard deviation estimates.

Many other models are saved in results/models/.
Modify the paths given in example_nn.sh as arguments to example_nn.py to test them, too!

We don't attempt to set up a GPU; you can ignore the warning WARNING:absl:No GPU/TPU found, falling back to CPU....

We include plots of cross-sections and Q split by hour in a notebook.

The split-Q plot uses data shared on Zenodo.
To run it yourself, download truth-jet-rot* from https://zenodo.org/record/6822267 and unzip them at a common path.
In the notebook, update the line DATAPATH = "..." in cell 8 to point to these unzipped data.

To launch it:

make env_nn/bin/activate
source env_nn/bin/activate
jupyter notebook
# in the notebook web interface, open: example_rotated_pv_msme.ipynb

Follow instructions in delphes/README.md for environment setup and execution. (cd delphes)