stukenov/sozkz-corpus-clean-enkk-fineweb-edu-v1

SozKZ Corpus Clean EN→KK (FineWeb-Edu) v1

Machine-translated parallel corpus of educational web texts (English → Kazakh), derived from FineWeb-Edu.

Overview

Source	FineWeb-Edu Score-2 (first 20M rows)
Languages	English (en) → Kazakh (kk)
Translation	CTranslate2, NLLB-200 (en→kk), greedy decoding
Filtering	Length (50–10K chars), exact dedup (xxhash), language detection (fasttext)
Rows	~18M (after filtering ~90% pass rate)
Format	Expandable Parquet shards (train-XXXXX.parquet)

Dataset Structure

Each row contains:

Field	Type	Description
text_en	string	Original English text from FineWeb-Edu
text_kk	string	Machine-translated Kazakh text
id	string	Original document ID from FineWeb-Edu
num_sentences	int	Number of sentences in the document

Usage

from datasets import load_dataset

ds = load_dataset("stukenov/sozkz-corpus-clean-enkk-fineweb-edu-v1", split="train")
print(ds[0])
# {'text_en': '...', 'text_kk': '...', 'id': '...', 'num_sentences': 5}

Processing Pipeline

1. Stream rows from HuggingFaceFW/fineweb-edu-score-2
2. Filter with cascade: length bounds → exact dedup (xxhash64) → language detection (fasttext lid.176)
3. Split text into sentences (regex-based)
4. Translate via CTranslate2 with NLLB-200 (en→kk), float16, greedy (beam=1), max 128 input / 200 output tokens
5. Upload each 1M-row chunk as a Parquet shard

Translation runs on 2× NVIDIA A10 GPUs in parallel.

Shard Layout

Shards are numbered sequentially (train-00000.parquet, train-00001.parquet, ...) and can be extended without renaming existing files.

Shards	Source
train-00000	Existing 902K pre-translated rows
train-00001 – train-00009	FineWeb-Edu rows 1M–10M (filtered)
train-00010 – train-00019	FineWeb-Edu rows 10M–20M (filtered)

Limitations

• Machine translation quality (NLLB-200) — not human-verified
• Sentence splitting is regex-based and may introduce segmentation errors
• Educational domain bias inherited from FineWeb-Edu
• Some truncation at 128 input tokens for very long sentences

License

CC-BY-4.0 (following FineWeb-Edu licensing)

Citation

@dataset{sozkz_corpus_clean_enkk_fineweb_edu_v1,
  title={SozKZ Corpus Clean EN-KK (FineWeb-Edu) v1},
  author={Saken Tukenov},
  year={2026},
  url={https://huggingface.co/datasets/stukenov/sozkz-corpus-clean-enkk-fineweb-edu-v1},
  note={Machine-translated from FineWeb-Edu using NLLB-200}
}