Dra. Mara Chinea Ríos

Publicaciones

• IberBench: LLM Evaluation on Iberian Languages

  IberBench is an extensible multilingual benchmark for evaluating Large Language Models on fundamental and industry-relevant NLP tasks across languages and varieties spoken in the Iberian Peninsula and Ibero-America. It integrates 101 datasets covering 22 task categories and evaluates 23 LLMs. Results reveal performance gaps across languages and tasks. IberBench provides an open-source evaluation pipeline and a public leaderboard.

  Ver publicación

• AI-Q: a Framework to Measure the Robustness of Text Classification Models

  AI-Q is a Python framework for evaluating and improving the robustness of text classification models through adversarial attacks and data augmentation. It provides 24 perturbation methods and robustness evaluation metrics, supporting both custom and HuggingFace models. AI-Q also enables dataset augmentation by reusing adversarial techniques, while experimental and human evaluations demonstrate its potential for improving model robustness.

  Ver publicación

• Profiling Cryptocurrency Influencers with Few-shot Learning

  This paper presents the PAN2023 Author Profiling shared task, focused on identifying cryptocurrency influencers on social media from a low-resource perspective. The task includes three subtasks addressing influencer profiling, interests, and intent using English Twitter data. A total of 27 teams participated, with their approaches evaluated against four baselines and the results analyzed and discussed.

  Ver publicación