edisimon/armenian-clean-text

Armenian Clean Corpus (pretraining + SFT bundle)

Combined, deduplicated, cleaned Armenian text assembled for pretraining and supervised fine-tuning of small language models. Built via the pipeline at https://github.com/EdikSimonian/armenian-gpt:

python 1_download.py        # fetch sources
python 2_prepare.py         # clean + dedup + merge
python 1_download.py --upload   # push this bundle

Contents

corpus/clean_text.txt.zst    zstd-compressed merged corpus
finetune/*.json              SFT Q&A files (ArmBench, Aya hy slice,
                             optional Qwen/Claude generated pairs)

Corpus statistics

• Uncompressed size: 62.85 GB
• Compressed size (zstd L19): 12.07 GB
• Paragraphs: 20,420,721
• Sources: 12 (see list above)

Deduplicated at the paragraph level across all sources. Quality-priority dedup order: Wikimedia → ARLIS → news → HPLT 3.0 → CulturaX → FineTranslations → mC4 → CC-100.

Fine-tuning Q&A files

• finetune/armbench_eval.json — 150 pairs
• finetune/armbench_train.json — 304 pairs
• finetune/armenian_qa_qwen.json — 1,104 pairs
• finetune/armenian_qa_qwen_short.json — 779 pairs
• finetune/aya_armenian.json — 27,069 pairs

Sources used

• Armenian Wikipedia (hywiki) — CC BY-SA 4.0
  https://dumps.wikimedia.org/hywiki/
• Armenian Wikisource (hywikisource) — CC BY-SA 4.0
  https://dumps.wikimedia.org/hywikisource/
• Armenian Wiktionary (hywiktionary) — CC BY-SA 4.0
  https://dumps.wikimedia.org/hywiktionary/
• Armenian Wikiquote (hywikiquote) — CC BY-SA 4.0
  https://dumps.wikimedia.org/hywikiquote/
• CC-100 Armenian — Common Crawl Terms of Use
  https://data.statmt.org/cc-100/hy.txt.xz
• HPLT 3.0 Armenian (hye_Armn) — CC0 1.0
  https://data.hplt-project.org/three/sorted/hye_Armn.map
• CulturaX Armenian (uonlp/CulturaX) — ODC-By 1.0
  https://huggingface.co/datasets/uonlp/CulturaX
• mC4 Armenian (allenai/c4) — ODC-By 1.0
  https://huggingface.co/datasets/allenai/c4
• Glot500 Armenian (cis-lmu/Glot500) — Research use (mixed)
  https://huggingface.co/datasets/cis-lmu/Glot500
• ARLIS Armenian legislation database — Public domain (HY gov)
  https://data.opendata.am/dataset/arlis-db
• Stanford CC-News (hy filter) — Common Crawl Terms of Use
  https://huggingface.co/datasets/stanford-oval/ccnews
• FineTranslations Armenian (hye_Armn) — ODC-By 1.0
  https://huggingface.co/datasets/HuggingFaceFW/finetranslations

License and attribution

This dataset is assembled from multiple publicly available sources for text-and-data-mining research on low-resource language modeling. Wikimedia sources (Wikipedia / Wikisource / Wiktionary / Wikiquote) are CC BY-SA 4.0 — strict application of the ShareAlike clause would require this derivative work to also be CC BY-SA 4.0. The bundle is published under a mixed-source-attribution notice invoking the EU TDM exception (Directive 2019/790 Art. 4) and US fair use for ML research.

Users are responsible for compliance with source license terms in their own jurisdiction. If you redistribute or fine-tune models on this data, please also preserve attribution to the upstream sources listed above.

Quick start

Download the compressed corpus:

from huggingface_hub import hf_hub_download
import zstandard as zstd

path = hf_hub_download(
    repo_id="edisimon/armenian-clean-text",
    filename="corpus/clean_text.txt.zst",
    repo_type="dataset",
)

# Decompress
with open(path, "rb") as fin, open("clean_text.txt", "wb") as fout:
    zstd.ZstdDecompressor().copy_stream(fin, fout)

Or reproduce via the pipeline:

git clone https://github.com/EdikSimonian/armenian-gpt.git
cd armenian-gpt
pip install -r requirements.txt
python 1_download.py --download   # uses this HF dataset as source