|
|
--- |
|
|
license: other |
|
|
task_categories: |
|
|
- text-generation |
|
|
language: |
|
|
- ru |
|
|
tags: |
|
|
- russian |
|
|
- wikipedia |
|
|
- ruwiki |
|
|
pretty_name: RuWikiBench Data |
|
|
size_categories: |
|
|
- 1K<n<10K |
|
|
--- |
|
|
|
|
|
# RuWikiBench |
|
|
|
|
|
RuWikiBench is a Russian benchmark dataset designed to evaluate analytical capabilities of large language models via Wikipedia-style article generation. |
|
|
This repository contains the data used by the RuWikiBench pipeline: extracted source texts and link mappings. |
|
|
|
|
|
|
|
|
## Data collection |
|
|
|
|
|
The dataset is based on a manually selected subset of articles from the Russian internet encyclopedia **"Ruwiki"** (Рувики). The selection targets diverse topics with sufficient external references. |
|
|
|
|
|
Collection procedure: |
|
|
|
|
|
1. **Article selection**: articles were manually chosen to cover diverse domains and to contain many external references. |
|
|
2. **Source downloading**: external sources referenced by each article were automatically downloaded (e.g., using `newspaper3k`). |
|
|
3. **Chunking into snippets**: source texts were split into fragments of approximately **~600 words** to emulate retrieval-augmented generation conditions. |
|
|
|
|
|
Below is the schematic breakdown of an article `https://ru.ruwiki.ru/wiki/Python`. |
|
|
<img src='article_entities.png' alt='Main article entities' width="60%"> |
