Software

This section showcases our open-source software tools and models, designed to advance research and applications in Natural Language Processing (NLP), Human-Robot Interaction, ethical AI, and multimodal learning. Our software includes instruction-tuned large language models, kernel-based learning platforms, and state-of-the-art architectures for grounded understanding, fact verification, and ethics-aware decision-making. Each tool is released with comprehensive documentation and is available on GitHub to support transparent, reproducible, and adaptable AI development for both academic and industrial applications.

ExtremITA – Italian Multi-Task Instruction-Tuned LLM

GitHub: https://github.com/crux82/ExtremITA

Description: ExtremITA is an instruction-tuned large language model built specifically for Italian. Based on the LLaMA architecture, it leverages instruction-following paradigms to achieve top performance in 22 diverse NLP tasks (1st place in 41% of tasks) in EVALITA 2023. ExtremITA uses LoRA adapters and the PEFT library for sustainable, modular fine-tuning, making it suitable for robust, multitask AI applications in Italian.

GrUT – Grounded Language Understanding for Robots

GitHub: https://github.com/crux82/grut

Description: GrUT (Grounded language Understanding via Transformers) is a neural framework designed for Human-Robot Interaction. It interprets natural language commands given to robots, grounding these commands in real-world entities and their properties. GrUT generates fully semantic representations (Frame Semantics) directly linked to the robot’s internal knowledge, significantly improving robotic understanding of complex spoken commands.

GAN-BERT – Generative Adversarial Networks for Few-shot Text Classification

GitHub: https://github.com/crux82/ganbert-pytorch

Description: GAN-BERT applies Generative Adversarial Networks (GANs) to transformer-based architectures (e.g., BERT) to enhance text classification performance, especially in few-shot scenarios. It reduces dependency on labeled data, enabling robust semantic classification even with limited training samples. GAN-BERT has been validated for its effectiveness in text categorization tasks across various domains.

MT-GAN-BERT – Multi-Task GAN for Robust Text Classification

GitHub: https://github.com/crux82/mt-ganbert

Description: MT-GAN-BERT extends GAN-BERT by integrating multitask learning capabilities. It simultaneously learns from multiple classification tasks, efficiently transferring knowledge across tasks and reducing overfitting. This approach ensures stable and robust performance, particularly beneficial for scenarios with scarce training resources across different linguistic tasks.

KeLP – Kernel-based Learning Platform for NLP

GitHub: https://github.com/SAG-KeLP/kelp-full

Website: http://www.kelp-ml.org

Description: KeLP is a Java-based machine learning platform specifically designed to facilitate kernel-based methods for NLP applications. By decoupling learning algorithms from kernel functions, KeLP enables flexible and efficient experimentation with structured data (trees, sequences, etc.), scaling up to large-scale language learning problems. It supports research and rapid prototyping in complex semantic tasks such as question answering, semantic role labeling, and more.

U-DepPLLaMA – Universal Dependency Parsing via Large Language Models

GitHub: https://github.com/crux82/u-deppllama

Description: U-DepPLLaMA introduces a novel method for universal dependency parsing using auto-regressive large language models (LLaMA2). It converts dependency parsing into a sequence-to-sequence problem, achieving state-of-the-art accuracy across 26 languages without relying on specialized parsing architectures. This project demonstrates how LLMs can simplify complex syntactic parsing tasks effectively.

MM-IGLU – Multi-Modal Interactive Grounded Language Understanding

GitHub: https://github.com/crux82/MM-IGLU

Description: MM-IGLU is a multimodal benchmark and model suite for grounded language understanding in 3D environments. It simulates interactions where an AI agent must execute or clarify natural language commands based on a visual world. The project includes both a text-only BART-based model and a multimodal model combining CLIP with LLaMA2-Chat-13B via LLaVA. MM-IGLU enables training and evaluation of models on command executability and clarification generation, setting a foundation for interactive agents in virtual or robotic environments.

MM-IGLU-IT – Italian Multi-Modal Grounded Language Understanding

GitHub: https://github.com/crux82/MM-IGLU-IT

Description: MM-IGLU-IT extends the MM-IGLU framework to Italian, offering the first large-scale dataset and models for multimodal grounded instruction following in Italian. Built from DeepL-translated and manually validated data, it supports classification and generation tasks in 3D environments. Leveraging LLaVA and LLaMA2-Chat-13B fine-tuned for Italian, MM-IGLU-IT achieves strong performance in both command feasibility assessment and interactive response generation, advancing multimodal LLM capabilities in non-English contexts.

FEVER-it – Fact Verification for Italian Texts

GitHub: https://github.com/crux82/FEVER-it

Description: FEVER-it provides the codebase for building and evaluating fact verification models in Italian, inspired by the original English FEVER benchmark. The software implements data preprocessing pipelines, retrieval modules for evidence from Wikipedia, and Natural Language Inference (NLI) models for claim verification. It serves as a full-stack prototype to develop AI systems capable of supporting or refuting claims using automated evidence retrieval and reasoning in Italian. The architecture is modular and designed for adaptation to other languages or domains.

EthicalNN – Neural Architecture for Ethics-Aware Decision Making

GitHub: https://github.com/crux82/nn-ebd

Description: EthicalNN is a PyTorch-based framework designed to integrate ethical reasoning into machine learning models. It combines neural network decision-making with formal ethical principles, enabling the development of systems that are both functionally effective and ethically sound. The architecture employs “truth-makers”—explicit rules representing ethical considerations—to guide learning processes, ensuring decisions align with predefined ethical standards. EthicalNN has been applied to various datasets, including COMPAS, Adult, German Credit, Default Credit Card, and Law School, demonstrating its versatility in addressing fairness and bias in AI systems.