gimmy256/afritour-bench-v1

AfriTour-Bench v1

AfriTour-Bench is a comprehensive evaluation benchmark designed to assess the accuracy, cultural sensitivity, and practical reliability of Large Language Models (LLMs) in the context of African tourism and travel.

It was developed as part of the sovereign AI stack for the East African region.

Dataset Description

The benchmark contains high-quality, grounded Question-Answer pairs derived from Wikivoyage articles covering all 54 African countries and major tourist destinations (e.g., Serengeti, Victoria Falls, Marrakech).

Unlike generic global benchmarks, AfriTour-Bench specifically probes an AI's ability to handle regional nuances, avoiding hallucinations about geography, respecting local customs, and providing safe practical advice for the continent.

• Language(s) (NLP): English
• License: CC-BY-SA 4.0 (Inherited from Wikivoyage)

Dataset Structure

The dataset is split into two primary subsets:

• benchmark_qa.jsonl (20%): The holdout evaluation set. Use this strictly for model benchmarking and comparing performance.
• training_qa.jsonl (80%): The training set. Can be used for Supervised Fine-Tuning (SFT) or Retrieval-Augmented Generation (RAG) conditioning.

Additionally, the pairs are categorized for targeted evaluation:

• cultural_qa.jsonl: Questions regarding customs, etiquette, religion, and dress codes.
• practical_qa.jsonl: Questions about visas, currency, transport, and border crossings.

Data Instances

A typical instance looks like this:

{
  "id": "uganda-0001",
  "source_title": "Uganda",
  "source_url": "https://en.wikivoyage.org/wiki/Uganda",
  "region": "East Africa",
  "type": "country",
  "category": "cultural",
  "difficulty": "medium",
  "question": "What cultural customs should visitors respect when visiting rural areas in Uganda?",
  "answer": "Visitors to rural Uganda should dress modestly, keeping shoulders and knees covered. It is customary to greet elders first and use the right hand for giving or receiving items.",
  "ground_truth_source": "wikivoyage",
  "language": "en"
}

Data Fields

• id: Unique identifier for the QA pair.
• source_title: The Wikivoyage article title the pair was generated from.
• source_url: Link to the original article.
• region: The African sub-region (e.g., East Africa, West Africa).
• type: Destination type (country, city_or_town, park_or_natural_site, region).
• category: The evaluation category (cultural, practical, wildlife, itinerary, safety).
• difficulty: Assessed difficulty of the question (easy, medium, hard).
• question: A natural, traveler-oriented question.
• answer: A highly accurate, grounded answer based purely on the source text.
• ground_truth_source: Identifies the origin of the facts (wikivoyage).
• language: Two-letter language code (en).

Generation Process

1. Source Collection: African destination articles were scraped from Wikivoyage using a polite crawler.
2. Filtration: Articles with fewer than 80 words were discarded.
3. Ground Truth Generation: A strict generation prompt was passed to Gemini 2.5 Flash, feeding it the raw text of the article and instructing it to generate {N} QA pairs across specific categories without utilizing any external parametric knowledge.

Intended Use

This dataset should be used to:

1. Benchmark Models: Evaluate foundational models (e.g., Llama 3, Gemma, Qwen) on their inherent knowledge of African geography and culture.
2. Prevent Harm: Specifically evaluate models on the safety and practical categories to ensure AI systems do not give dangerous travel advice.
3. Train Regional AI: Fine-tune models to adopt the persona of a highly knowledgeable local African travel concierge.