---
task_categories:
- question-answering
- text-generation
language:
- el
tags:
- wikipedia
- greek
- biography
- rag
- gemma
- synthetic-dataset
size_categories:
- 100K<n<1M
pretty_name: Greek Biography QA Dataset
dataset_info:
- config_name: full_dataset
  features:
  - name: title
    dtype: string
  - name: full_text
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 617649777
    num_examples: 142200
  download_size: 104499197
  dataset_size: 617649777
- config_name: subset
  features:
  - name: title
    dtype: string
  - name: full_text
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_bytes: 18181659
    num_examples: 1650
  download_size: 2190731
  dataset_size: 18181659
configs:
- config_name: full_dataset
  data_files:
  - split: train
    path: full_dataset/train-*
- config_name: subset
  data_files:
  - split: train
    path: subset/train-*
---
 
# Wiki_Greek_Personalities Dataset

Developed by **NCSR-D**, this dedicated dataset focuses on expanding LLM knowledge regarding Greek public figures, enhancing domain knowledge, and providing accurate biographical data on prominent Greek personalities. The dataset systematically transforms unstructured biographical data derived from a recent dump of the Greek Wikipedia into highly structured training examples. The full dataset contains 142,200 Question-Answering (QA) pairs for 16,075 Greek public figures, while the subset contains 1,650 QA pairs for 170 Greek public figures.
 
## Dataset Structure
 
The dataset is provided in a standard JSON format containing a list of objects. Each object represents a single QA context entry:
 
## Schema Overview
 
Both configurations share an identical schema containing the following fields:
 
* `title` (string): Target subject's identity name.
* `full_text` (string): The cleaned reference source text snippet from Greek Wikipedia.
* `question` (string): Natural language prompt querying essential biographical facts.
* `answer` (string): The validated accurate response text.
 
### Data Instance Example
 
```json
{
  "title": "Person",
  "full_text": "Ο Person είναι Έλληνας πολιτικός ...",
  "question": "Πότε και πού γεννήθηκε ο Person;",
  "answer": "Ο Person γεννήθηκε στις Date στον Location."
}
```

## Dataset Configurations & Statistics
 
The repository currently features two distinct configurations built dynamically through an advanced LLM generation and validation pipeline:
 
### 1. `full_dataset`
* **Size:** 142,200 validated QA pairs.
* **Coverage:** 16,075 unique Greek public figures.
* **Domains Covered:** Politics, science, academia, culture, arts, sports, business, and public life.
 
### 2. `subset`
* **Size:** 1,650 validated QA pairs.
* **Coverage:** 170 unique Greek public figures.
* **Purpose:** A smaller version of the full dataset focused exclusively on the most recognizable public figures, ideal for rapid evaluation, validation, or targeted benchmarking.

## Dataset Creation & Curation Pipeline
 
### 1. Extraction & Scraping
NCSR-D identified and targeted **16,075 unique Wikipedia pages** matching Greek individuals. The system extracted specific Wikipedia person properties directly from the latest dump to ensure up-to-date and robust context mappings.
 
### 2. Synthetic QA Generation
The language model `google/gemma-3-12b-it` (Gemma Team, 2025) was used to automatically draft QA pairs. The model was constrained to generate questions and answers *exclusively* from the biographical details embedded within each individual's text profile to combat hallucinations.
 
### 3. Dual-LLM Quality Control & Filtering
To ensure clean data and exceptional linguistic consistency, a second-pass evaluation was executed using `google/gemma-4-12B` (Gemma Team, 2026). The evaluation system filtered out data based on strict validation parameters:
* **Factual correctness** and total consistency against the source text.
* **Linguistic compliance**, measuring grammar, syntax, phrasing clarity, and overall Greek structural correctness.
 
Pairs failing any parameter were pruned from the final training split.

## Loading and Usage
 
Load either configuration natively using the Hugging Face `datasets` package:
 
```python
from datasets import load_dataset
 
# To load the full collection of 142,200 QA pairs:
dataset_full = load_dataset("PharosAIFactory/Wiki_Greek_Personalities", "full_dataset")
 
# To load the smaller validation subset of 1,650 QA pairs:
dataset_subset = load_dataset("PharosAIFactory/Wiki_Greek_Personalities", "subset")
```
 
## Additional Information
* **Curator Organization:** NCSR-D / [PharosAIFactory](https://huggingface.co)