The UC-UCO-Plenitas team participated in the PRESTA 2025 challenge, focused on question answering over tabular data in Spanish using the DataBenchSPA benchmark. This benchmark is the first of its kind in the Spanish language and includes real-world tables with diverse data types, designed to evaluate system capabilities in answering natural language questions. The team implemented a solution leveraging GPT-4o, a multimodal large language model developed by OpenAI, known for its real-time, multi-input processing capabilities. GPT-4o was used to handle text-based question answering tasks, with the final system developed using under 150 lines of code, integrating the evaluation functions provided by the organizers. Among 23 competing teams, the UC-UCOPlenitas team secured 7th place, achieving 66.0% accuracy, showcasing the model's potential and competitive performance against other state-of-the-art approaches. While not reaching the top three, the team's results highlight opportunities for further performance optimization through better prompt design and fine-tuning. The paper also provides insights into deep learning architectures, particularly transformers, and emphasizes the role of large language models (LLMs) in advancing natural language understanding over structured datasets.
Lexical complexity relates to complexity of words. Its assessment can be beneficial in a number of
ifelds, ranging from education to communication. For instance, lexical complexity studies can assist in
providing language learners with learning materials suitable for their proficiency level or aid in text
simplification [
Dataset The IberLEF 2025 shared task is centered on Question Answering over Tabular Data, making
use of the newly developed DataBenchSPA benchmark [
Deep Learning Deep learning is a type of machine learning that uses artificial neural networks
with many layers (hence "deep") to model complex patterns in data. It’s especially good at handling
unstructured data like images, text, and audio. Deep learning enables computers to learn from data
much like the human brain does. Instead of being programmed with specific rules, a deep learning
system figures out the rules on its own by training on large datasets[
Transformers Transformers are a type of deep learning architecture introduced by [
Large Language Models (LLMs) Large Language Models are deep learning models trained on
massive amounts of text data to understand, generate, and manipulate human language. These models
use neural networks—typically transformer architectures—to learn patterns in text and perform tasks
like Text generation, Translation, Summarization, Sentiment analysis, Question answering and Code
generation. A Large Language Model is a type of artificial intelligence (AI) trained to understand,
generate, and interact using human language. These models are built using deep learning techniques
(usually transformers) and are trained on vast amounts of text data — including books, websites, and
documents — to learn grammar, facts, reasoning, and context[
GPT4o GPT is a type of Large Language Model (LLM) developed by OpenAI [
• Solve math problems from handwritten notes or photos.
Setup and functionality The setup and functionality of an automated system designed to answer natural language questions about tabular datasets, in the context of the PRESTA 2025 (IberLEF) competition. The system uses OpenAI’s GPT-4o language model to generate Python code that directly answers each question by manipulating the data using the pandas and numpy libraries. Its workflow is structured into four main stages: prompt generation, model query, code execution, and answer evaluation.
To get started, Python 3.8 or higher is required, along with the installation of key dependencies: pandas, numpy, openai, nest_asyncio, and the databench_eval library, the latter provided as part of the oficial evaluation setup. Once the environment is ready, a valid OpenAI API key must be configured. Although the base code defines it as a constant, best practice is to store the key in an environment variable using export OPEN_AI_KEY="sk-..." and access it within the script using os.getenv("OPEN_AI_KEY"). It is also recommended to download the datasets in Parquet format and store them under the path /̇databenchSPAdatasetall.parquet, although HuggingFace datasets can also be used via the "hf:// prefix".
The system loads the development question set (qa_dev) using the load_qa function. For each question, it dynamically constructs a prompt that instructs the model to generate a single line of code within a predefined answer(df) function. This line should directly return the answer to the question using only the provided DataFrame. For example, for the question "What is the average number of bedrooms?" on the airbnb dataset, the system generates a prompt like:
"You are a pandas code generator. Your goal is to complete the function provided... def answer(df: pd.DataFrame): return".
The model might respond with a line such as df[’bedrooms’].mean(), which the system executes dynamically using exec() to compute the final answer. This technique allows the automatic evaluation of the model’s reasoning capabilities on tabular data. The generated results are saved in a file named predictions.txt, which can then be submitted as an oficial run. Additionally, performance is measured by comparing the generated responses to ground truth labels using the Evaluator module. This baseline system achieves approximately 49% accuracy on the development set.
An optional utility function, column_generator(), is also included. It generates prompts to filter and select only the most relevant columns needed to answer each question. This is useful for reducing input size when dealing with models that have context limitations. Despite its efectiveness, this system carries potential security risks due to the use of exec() for code execution, and should only be used in controlled environments.
Evaluation An automated evaluation function is currently provided to handle most of the assessment process. When a participant uploads a submission, the default evaluation function from the databench_eval package is executed, comparing the submission against the ground truth set. This function has been modified to be less strict than the one used in the initial experiment. The adjustment accommodates slight variations in formatting, allowing smaller models to avoid penalties for minor errors. Given the heuristic nature of the evaluation, the characteristics of the models being used, and the open-ended nature of the task, the organizers will manually review the top-scoring submissions before selecting a winner.
Types of Answers Expected According to the expected answer types: • Boolean: Valid answers include True/False, Y/N, Yes/No (all case insensitive). • Category: A value from a cell (or a substring of a cell) in the dataset. • Number: A numerical value from a cell in the dataset, which may represent a computed statistic (e.g., average, maximum, minimum). • List[category]: A list containing a fixed number of categories. The expected format is: "[’cat’, ’dog’]". Pay attention to the wording of the question to determine if uniqueness is required or if repeated values are allowed.
• List[number]: Similar to List[category], but with numbers as its elements.
Competition phase In the competition, our team employed GPT-4o as the primary method for solving the tasks, achieving the following results: In the competition phase, out of 23 total participants, the UCUCO-Plenitas team secured 7th place with an accuracy of 66.0% . This result reflects a solid performance considering the number of competitors, although there is clear room for improvement to reach the top positions. The top three teams — itunlp , sonrobok4 , and hcerezo — achieved significantly higher accuracies (85–87%), indicating a high level of performance in the task. Teams ranked from 4th to 6th (e.g., LyS Group , quang3010 , and ScottyPoseidon ) also outperformed UC-UCO-Plenitas by a noticeable margin. Despite not reaching the podium, UC-UCO-Plenitas demonstrated competitive capability, especially when compared to other mid-to-lower-ranked teams. With further refinement, particularly in precision and fine-tuning of classification strategies, the model could potentially move up in future rankings. See table 1.
The participation of the UC-UCO-Plenitas team in the PRESTA 2025 competition demonstrated the growing applicability of Large Language Models, particularly GPT-4o, in the domain of question answering over tabular data. Achieving a 66.0% accuracy rate, the team’s performance places them in the top third of participating systems and highlights the feasibility of using GPT-4o in data-intensive Spanish-language NLP tasks. The results validate the use of minimal coding approaches with powerful pre-trained models and point toward the potential of improved performance through enhanced prompt engineering and fine-tuning strategies. The competition served as a valuable benchmark for evaluating the capabilities of modern LLMs in multilingual and structured data contexts. Future eforts will focus on improving system generalization and interpretability to reach the performance levels of the top-ranking teams.
During the preparation of this work, the author(s) used GPT-4 and Grammarly in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.
The results are available via • PRESTA Codabench, • PRESTA Dataset.