Welcome to the HSF Jet Classification Challenge 2026!
Organized by the High Energy Physics Software Foundation (HSF) in collaboration with the hls4ml community, this competition introduces participants to one of the most important machine learning applications in modern particle physics: jet classification.
When protons collide at extremely high energies inside the Large Hadron Collider (LHC), they produce unstable particles that rapidly decay into collimated sprays of hadrons known as jets. Determining the particle that produced a jet is a fundamental task in experimental particle physics and plays a crucial role in precision measurements and searches for new physics.
In this challenge, you will develop a machine learning model that classifies jets into one of five categories:
Rather than working directly with detector images or low-level particle information, you will use 16 high-level physics observables that capture the internal structure of each jet. Your goal is to predict the probability that each jet belongs to each of the five particle classes.
This competition is designed as an educational introduction to machine learning for High Energy Physics (HEP). Whether you are a student beginning your ML journey or an experienced practitioner interested in scientific applications, you will gain hands-on experience with a realistic particle physics classification problem while learning how modern machine learning techniques are applied at the LHC.
To help participants get started, the competition includes an official starter notebook demonstrating a complete baseline workflow—from data loading and preprocessing to model training, evaluation, and submission generation. Participants are encouraged to experiment, compare approaches, and improve upon the provided baseline while maintaining reproducible and well-documented machine learning workflows.
Submissions are evaluated on Multiclass Log Loss (also known as Categorical Cross-Entropy) between the predicted probabilities and the observed target classes. Lower Log Loss values indicate better model performance. For each jet ID in the test set, you must predict the probability for each of the five target classes (g, q, t, w, z). The submission file must contain a header and follow the specified CSV format.