Code & Reproducibility

Packages:

qJAX: Flexible Machine Learning with Tsallis statistics, built on JAX

Most machine learning tools quietly assume a single shape for randomness and uncertainty: the bell curve, softmax attention, Shannon entropy. These are baked in as fixed rules, and real data often doesn't follow them. Labels are noisy, outcomes have heavy tails, and attention sometimes needs to focus sharply on a few inputs instead of spreading weight over everything. qjax replaces that fixed rule with a single tunable parameter, q, so these familiar tools can stretch to fit the data instead of forcing the data to fit them. Because q is fully differentiable, a model can learn the right setting during training instead of you having to guess it. In practice this means classifiers that resist memorizing mislabeled data, distributions that capture heavy tails a standard Gaussian would miss, and attention that can be as sparse or as dense as the problem needs, all as drop-in JAX functions compatible with jit, vmap, and the training code you already have.

GraphNetz: Statistically rigorous Graph Learning benchmarking

Whether you are proposing a new GNN architecture, testing a model on a new graph domain, or comparing existing methods across graph types, GraphNetz turns the usual “train, evaluate, table of accuracies” workflow into a reproducible statistical report. Instead of reporting point estimates alone, it provides confidence intervals for each result, paired model comparisons with multiple-testing correction, and rank-based summaries across datasets using critical-difference diagrams. The goal is not just to crown a leaderboard winner, but to give researchers a principled way to quantify uncertainty, compare methods fairly, and produce the exact evidence reviewers often ask for in graph-learning papers. Detailed methodology in Da Costa and Modenesi (2026) <doi.org/10.48550/arXiv.2605.09099>

epiworldRcalibrate: Fast and Effortless Calibration of Agent-Based Models using Machine Learning (GitHub)

Provides tools and pre-trained Machine Learning [ML] models for calibration of Agent-Based Models [ABMs] built with the R package 'epiworldR'. Implements methods described in Najafzadehkhoei, Vega Yon, Meyer and Modenesi (2025) <doi:10.48550/arXiv.2509.07013>. Users can automatically calibrate ABMs in seconds with pre-trained ML models, effectively focusing on simulation rather than calibration. Bridges a gap by allowing public health practitioners to run their own ABMs without the advanced technical expertise often required by calibration.