One collision, one place in the dataset

My first useful unit of data was a collision. One collision can leave several short segments in the muon spectrometer. They are clues from the same event, so I keep them together from the first read of the ROOT file through to the final test.

Give the collision an identity

Event numbers can repeat across independently generated samples. I pair the run and event numbers with the identity of the simulation sample. Copies or reprocessings of the same collision need to keep the same identity. The split uses the collision identity and a fixed seed, so rereading the events in a different order gives the same assignments. I save those assignments alongside the information needed to trace each extracted record back to its source.

Imagine one collision leaves three segments. If two land in training and the third lands in testing, the model gets part of the test event during training. With an event-level split, all three go to the same partition. The training set fits the model, the validation set guides choices such as the score threshold, and the test set is where I check those choices on held-out collisions.

Keep the answer in a separate envelope

The model inputs describe reconstructed segments, starting with their positions and directions. In simulation I can also read the generated decay positions, which I keep in a separate truth record. An eligible target is a generated decay that meets the agreed detector region and selection. A complete record with zero eligible decays supplies a negative example. A missing truth record leaves the answer unknown.

I checked that extraction preserves the collision identity, the segment arrays line up and the split joins back to the same collisions. That gave me a way to test how the data move through the pipeline. The next step for physics training is to settle the eligible decay region and confirm the coordinate conventions. Two files that open perfectly can still disagree about where a metre starts.