Yacine Haddad

I just had a look, the aqora test command seems to evaluate predictions over the entire validation dataset, as the scoring logic expects outputs for all rows. Currently, I see no built-in option to restrict evaluation to a subset (e.g., the first N rows), which explains the error you’re encountering when using only part of the data.

That said, one idea is to split the training dataset into smaller subsamples, train on one part and validate on the other. Since the data originates from the same physics process, good performance on the local validation set should generalize well to the full validation dataset on the backend.

Yes, it’s definitely possible to use Google Colab instead of Visual Studio Code for the coding part. Colab is actually great for running Python code without any local setup, and it supports many popular libraries right out of the box. The only thing to keep in mind is that you’ll need to upload your prediction file here once you’re done.

Hi Marcin,

Yes, you’re absolutely right. Lower scores are better. The leaderboard score is based on the Kullback-Leibler (KL) divergence, which measures the difference between your predicted distributions (In our scase the Jet Multiplicity Distribution and Leading & Subleading Jet $p_T$) and the ground truth. A perfect score of 0 would mean your predictions exactly match the true distributions (i.e $P(i) = Q(i)$), though that is very hard to achieve in practice, so it's very unlikely.

No worries at all about the question, happy to clarify, and thanks for checking in.

Cheers, Yacine

Quantum computers are, in principle, a natural framework for simulating particle interactions and QFTs in HEP due to their ability to efficiently handle quantum states and entanglement. Traditional simulations of QFTs rely on lattice-based methods (like Lattice QCD) and perturbative methods, which require immense classical computational resources. Quantum computing can address these challenges by representing quantum states directly on qubits, evolving them using quantum circuits, and leveraging quantum entanglement to capture correlations more efficiently. While this is an active research area and we are not there yet! Many researchers are working on it, and interest in the field has grown significantly over the years. See Michele's answer for more information

Yes, partons and jets are related, as partons (quarks or gluons) from the hard process undergo hadronization, producing sprays of particles that form jets. The main challenge is establishing a mapping between initial partons and final-state jets. Commonly used features for analysis include transverse momentum ($p_T$), pseudorapidity ($\eta$), azimuthal angle ($\phi$), and the angular separation $\Delta R = \sqrt{\Delta \eta^2 + \Delta \phi^2}$, which helps associate partons with jets.

Additional useful features include jet energy and momentum fractions to compare jets with their originating partons. Matching techniques such as $\Delta R$-based nearest neighbor assignment or momentum fraction consistency are commonly used. Depending on the dataset, a simple kinematic-based matching may be sufficient, but more advanced techniques could be explored if needed.

Yes it is a general feature of the way the data is produced. In the current challenge we want to know how we can produce 2 jets and above. so at the end we don't evaluate the solution when jet are below 30 GeV. The KL divergence is calculated for the following bins :

pt_bins = [30, 40, 50, 60, 70, 80, 90, 110, 140, 180, 230]
n_jets_bins = [2, 3, 4, 5]

At the end, in your solution is it more important to model events with at least 2 jets.

That's seems about right. We are should see some events that are expected to fall beyond the detection phase-space. I think it is safe to consider neglecting these events at the moment. For instance you can required to have n_jets >= 1 so you can have at least one object in the final state.

The data is structured as follow, if no jet is found then then this means both jets are below 30 GeV threshold and this event can be discarded. It is possible to have this special case.

So you can basically make a selection to choose the right events to work with. For example, can do:

jets_p4  = jets_p4[jets_p4.pt>30] 
partons_p4 = partons_p4[jets_p4.pt>30]

How many events do you see that have zero's everywhere?