Data/Monte Carlo comparison framework for the ATLAS flavour tagger input variables.

The ATLAS Flavour-Tagging Algorithms sub-group (FTAG) is devoted to developing advanced jet tagging methods using state-of-the-art deep learning models. Among them, Graph Neural Networks (GNNs) are emerging as a promising approach due to their ability to exploit the well-structured anatomy of collision events. Recently, a new GNN-based tagger has been introduced for flavour tagging in ATLAS, which improves upon previous taggers by incorporating new input variables and achieving better performance. To ensure that the Monte Carlo (MC) datasets, on which the network is trained, agree with the Data it is necessary to quantify the Data/MC differences in the input variables. A new framework that allows the user to obtain quick Data/MC checks for the jet variables has been developed. The workflow starts from the event selection to the production of the Data/MC plots with the PUMA functionalities for a wide range of variables of interest. In this communication, I will present the framework's main features along with the results obtained.