---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: file_name
    dtype: string
  - name: id
    dtype: string
  - name: language
    dtype: string
  - name: language_score
    dtype: float64
  - name: source
    dtype: string
  - name: subsource
    dtype: string
  - name: text
    dtype: string
  - name: fineweb-nemotron-edu-score
    dtype: float32
  - name: fineweb-mixtral-edu-score
    dtype: float32
  - name: alignment-score
    dtype: float32
  - name: fasttext-quality-score
    dtype: float32
  - name: manipulative-score
    dtype: float32
  - name: gec-score
    dtype: float32
  - name: __index_level_0__
    dtype: int64
  - name: median-score
    dtype: float64
  - name: formula-score
    dtype: float64
  - name: fineweb-nemotron-edu-score-int
    dtype: int64
  - name: fineweb-mixtral-edu-score-int
    dtype: int64
  - name: alignment-score-int
    dtype: int64
  - name: fasttext-quality-score-int
    dtype: int64
  - name: manipulative-score-int
    dtype: int64
  - name: gec-score-int
    dtype: int64
  - name: max-score-int
    dtype: int64
  splits:
  - name: train
    num_bytes: 152867030506
    num_examples: 16702287
  download_size: 69984117765
  dataset_size: 152867030506
license: cc-by-4.0
task_categories:
- text-generation
language:
- uk
tags:
- ukrainian
- pretraining
pretty_name: Lapa High Quality Pretraining Dataset
---


# Dataset Card for Lapa High Quality Pretraining Dataset

## Dataset Description

**Dataset Summary**

This dataset is a high quality subset of pretraining corpus for Ukrainian language. It was filtered using 6 models, measuring different quality aspects of the data:
- lapa-llm/alignment-score-model - Alignment - filtering for disinformation
- lapa-llm/gec-score-model - Grammatical Correctness of the text
- lapa-llm/fineweb-nemotron-edu-score - Educational Value of the text
- lapa-llm/fineweb-mixtral-edu-score - Educational Value of the text
- lapa-llm/manipulative-score-model - How manipulative is the text
- lapa-llm/fasttext-quality-score - Text Coherence (how close is it to ELI5 from Reddit style of explanations)

All models are available in this collection: https://huggingface.co/collections/lapa-llm/lapa-v012-pretraining  

We perform CDF binning and perform max ensembling, then select data only from the highest performing bucket, which is this dataset.

Additional measure, `formula-score`, combines most of them into one metric:

```python
def formula_score(item):
     item["formula-score"] = np.median([item["fineweb-nemotron-edu-score"], item["fineweb-mixtral-edu-score"], item["fasttext-quality-score"],]) * item["alignment-score"] * item["manipulative-score"] * item["gec-score"]
     return item
```

This provides a balanced measure of quality across all classifiers.

**Languages**

- Ukrainian (uk)



## Dataset Creation

**Source Data**

- Base dataset: Kobza, FinePDFs, FineWeb, UberText




## Considerations for Using the Data

**Social Impact**

Aims to strengthen the Ukrainian-language LLM ecosystem and improve accessibility of language technology for Ukrainian speakers.



## Citation



**BibTeX**

TBD



## License

CC-BY-SA-4.0

---

*This dataset is part of the "Lapa" - Ukrainian LLM initiative to advance natural language processing for the Ukrainian language.*