This repository contains the training and inference code for our paper:
"Adapting Vision-Language Models for Neutrino Event Classification in High-Energy Physics" (submitted to Nature Communications).
We fine-tune the multimodal large language model LLaMA 3.2-11B Vision-Instruct on simulated DUNE Near Detector(LArTPC) pixel maps to classify neutrino interactions into three categories:
- NuE CC — Electron Neutrino Charged Current interaction
- NuMu CC — Muon Neutrino Charged Current interaction
- Neutral Current (NC)
The model takes as input 2D projections (xz, yz) of 3D detector events and learns to recognize distinctive features such as fuzzy electron showers (NuE CC) or long muon tracks (NuMu CC).
Code for running LLama 3.2 Vision and the CNN Baseline are provided in their respective directories, please follow instructions in their respective README.md files.