proxectonos/Galician-Generative-Eval-Prompts

Galician Generative Evaluation Prompts

Dataset Summary

This dataset contains a stratified sample of 60 input texts (prompts) for continuation and their respective original references, selected to cover a variety of linguistic registers and domains. It is designed to be executed within the Galician Rapid Human Evaluation Framework (galician-rapid-human-gen-eval) for LLMs in Galician.

Its primary purpose is to act as an input for the evaluation of model's open-ended generation capabilities during development, continual pre-training and fine-tuning. This allows researchers to audit language adherence, linguistic quality, and content quality of the generated text. It can be used also for the evaluation of final models to complement automatic evaluations.

All scripts for extracting model predictions, running the sampling and building the spreadsheets for annotators are maintained in the proxectonos GitHub repository. The Evaluation Guidelines for annotators are also maintained in the same repo.

Dataset Structure

Data Fields

Each instance in the dataset is structured in JSONL format and contains the following fields:

• id (string): Unique identifier of the original document.
• domain (string): The specific domain or linguistic register the text belongs to.
• prompt (string): The input text provided to the language model to generate a continuation.
• reference (string): The original continuation or reference text, provided as a supportive guide for the human annotator.

Domain Distribution

The dataset is balanced to test the model's versatility across different vocabularies and structural formats:

• legal: Administrative and legal texts (e.g., Official Gazette of Galicia).
• xornalistico: News and journalistic articles.
• enciclopedia: Factual and encyclopedic texts.
• cientifico: Academic abstracts and research papers.
• literario: Fiction and narrative texts.
• blog: Informal web texts and opinion pieces.
• dialogo: Conversational transcripts and subtitles.

Aknowledgements

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA. (Esta publicación del proyecto Desarrollo de Modelos ALIA está financiada por el Ministerio para la Transformación Digital y de la Función Pública y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU)