somosnlp-hackathon-2026/Onexe-QA-Dataset

Dataset Card: Canarian Linguistic Evaluation Dataset (QA without Answers)

Dataset Summary

This dataset has been designed specifically for evaluating the dialectal, linguistic, and cultural understanding of Large Language Models (LLMs) within the context of Canarian Spanish.

It contains 4,683 evaluation questions based on the official lexicon of the Academy of Canarian Language (Academia Canaria de la Lengua - ACL). Each record presents a linguistic query phrased using one of 41 distinct question templates generated randomly yet reproducibly. The main goal is to measure a model's ability to identify Canarian regional terms under diverse interrogative structures and evaluate its robustness to lexical and grammatical variations without having direct access to the answers in this subset.

Dataset Structure

The dataset is provided in CSV (qa_without_answer.csv) format with the following fields:

Field	Type	Description
question	string	The evaluation query formulated to test the Canarian term, varying across 41 different question styles.
country	string	The geographic or cultural anchor region for the dialectal term. All records: Canarias.
original_word	string	The original Canarian word/term as registered in the source dictionary.

Data Distribution

Geographic Breakdown (country)

Region / Territory	Records	Percentage
Canary Islands (Spain)	4,683	100%
Total	4,683	100%

Evaluation Classification

• Question Variability: Queries are uniformly distributed across 41 syntactic and dialectal interrogative styles to prevent fixed-prompt phrasing bias during benchmarking.

Key Use Cases

1. Dialectal Robustness Benchmarking: Verifies whether an LLM can understand questions asked in highly varied ways (e.g., "¿A qué se refiere concretamente la palabra...", "¿Cómo se interpreta...", "¿Qué quiere decir el término...") while keeping the core semantic query identical.
2. Canarian Lexicon Inference Testing (Zero-Shot / Few-Shot): Acts as a test set to present queries to an LLM under evaluation and measure whether its output aligns semantically with the official regional definitions.
3. Regional Alignment Profiling: Helps measure the model's linguistic sensitivity to Western Iberian-Atlantic variations (Canarian Spanish) compared to standard neutral Spanish.

Included Prompt Examples

Example 1: Direct Focused Interrogative Style

• Question: ¿A qué se refiere concretamente la palabra abacorante en Canarias?
• Original Word: abacorante
• Country: Canarias

Example 2: Scenario/Usage Interrogative Style (Common Speech)

• Question: Si alguien en Canarias usa la palabra abaifado, ¿qué quiere decir?
• Original Word: abaifado, da
• Country: Canarias

Example 3: Technical/Lexical Meaning Style

• Question: ¿Qué expresa la palabra canaria abade?
• Original Word: abade
• Country: Canarias

---

License

This dataset and its generation tools are distributed under the MIT License. You are free to copy, distribute, modify, and use it for commercial and private purposes, provided that the original copyright notice and license statement are included.