README.md

---
license: apache-2.0
dataset_info:
  features:
  - name: text
    dtype: string
  splits:
  - name: train
    num_examples: 142178930
  - name: validation
    num_examples: 71208
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
task_categories:
- text-generation
language:
- ru
size_categories:
- 100M<n<1B
---

# Cultura-Ru-Edu

The `Cultura-Ru-Edu` dataset consists of Russian educational web pages filtered from the [`uonlp/CulturaX`](https://huggingface.co./datasets/uonlp/CulturaX) dataset.

The dataset creation was inspired by [`HuggingFaceFW/fineweb-edu`](https://huggingface.co./datasets/HuggingFaceFW/fineweb-edu), but with a focus on the Russian language.
By filtering the dataset based on educational criteria, the `Cultura-Ru-Edu` dataset is both high-quality and large enough to train a Russian-focused language model for tasks requiring knowledge of the world.

## Dataset curation

To create this dataset, we annotated a subset with the `Meta-Llama-3-70B-Instruct` model, trained a classifier on it, and then applied it to the entire dataset, keeping only the high-quality samples.

### Annotation

Follow [`deepvk/cultura_ru_edu_llama3_annotations`](https://huggingface.co./datasets/deepvk/cultura_ru_edu_llama3_annotations) to see details about creating the annotation dataset.

### Training classifier

We trained a classifier based on the [`USER-base`](https://huggingface.co./deepvk/USER-base) model.
Unlike the original FineWeb-Edu pipeline, we used binary classification, where the positive class includes samples with a score of 3 and higher.
We found this approach more stable due to the high imbalance in the annotation dataset.

### Dataset scoring

We converted the classifier to ONNX format and applied it to the Russian part of the [`uonlp/CulturaX`](https://huggingface.co./datasets/uonlp/CulturaX) dataset.
The original dataset contained approximately 800 million documents, and after filtration, only 140 million documents remained (~17.5% of the original dataset).

## Dataset information

Each sample contains only one property — `text`, the original text document.

Some notes:
- This dataset is a filtered version of the larger, multilingual [`uonlp/CulturaX`](https://huggingface.co./datasets/uonlp/CulturaX) dataset. No other information was added or removed.
- Since the original dataset consists of parsed web pages, there may still be artifacts in the text header or footer. Future work may include detecting and removing such blocks.

## Usage

To use this dataset, one may simply use the `datasets` API.

```python
from datasets import load_dataset

cultura_ru_edu = load_dataset("deepvk/cultura_ru_edu", split="train", streaming=True)

```

Note that the dataset size is approximately 500GB, so it is better to use streaming or download it directly via Git LFS.

## Citations

```
@misc{deepvk2024cultura-ru-edu,
    title={Cultura-Ru-Edu},
    author={Spirin, Egor and  Sokolov, Andrey},
    url={https://huggingface.co./datasets/deepvk/cultura_ru_edu},
    publisher={Hugging Face}
    year={2024},
}
```