This dataset provides parallel translations of scientific and medical texts from Arabic (original) and English (source) into Russian, generated by two state‑of‑the‑art language models:

Gemma 3:4B (Google)
LLaMA 3.1:8B (Meta)

The corpus is built from four established Arabic–English corpora (see Sources below) and is intended for machine translation, model evaluation, and linguistic research.

Motivation and Model Selection

We conducted a pilot experiment on a random sample of 50 sentences (10 from each source) using six different models:

llama3.2:3b
gemma3:4b
phi3:mini
mistral:7b-instruct
llama3.1:8b
qwen2.5:7b

Expert evaluation (accuracy, terminology, grammar) showed that Gemma 3:4B and LLaMA 3.1:8B produced the most reliable translations for scientific and medical terminology.
Both models were therefore selected for the full‑scale translation of the entire corpus (52 384 original entries).

After filtering out entries where one of the translations was missing, we retained 52 335 fully aligned pairs (99.9% of the original).

Data Fields

Each record contains the following columns (no extra statistics are stored in the dataset itself):

Field	Type	Description
`arabic`	string	Original Arabic text
`english`	string	English source (from original corpus)
`ru_gemma3_4b`	string	Russian translation by Gemma 3:4B
`ru_llama3_1_8b`	string	Russian translation by LLaMA 3.1:8B
`source`	string	Origin of the text (corpus name)

All statistics (token counts, character counts, source distribution, translation similarity) are provided in the companion file dataset_stats.json.

Statistics

Overall

Total pairs: 52,335
Parquet chunks: 2

Sources

Source	Count	%
PEACH: a sentence-aligned Parallel English–Arabic Corpus for Healthcare	38,080	72.8%
MeSpEn_Glossaries (https://zenodo.org/records/1493393)	8,691	16.6%
Misraj/Tarjama-25	5,062	9.7%
NAMAA-Space/ASCAT-Arabic-Scientific-Translation	502	1.0%

Text length (average per pair)

Language / Model	Chars	Tokens
Arabic	103.6	17.03
Russian (Gemma)	127.7	16.27
Russian (LLaMA)	136.9	17.69

Translation Agreement (Gemma vs LLaMA)

Based on detailed analysis of the full corpus:

Exact matches: 3 263 (6.23%)
High similarity (>70% Jaccard): 19 835 (37.90%)
Low similarity (<70%): 32 500 (62.10%)

For full comparison statistics (length differences, word counts, similarity distribution) please refer to stats.txt and dataset_stats.json in this repository.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ArabicNLPWorld/arabic-russian-scientific-translations", split="train")

# Example
for example in dataset.select(range(3)):
    print(f"AR: {example['arabic']}")
    print(f"EN: {example['english']}")
    print(f"Gemma: {example['ru_gemma3_4b']}")
    print(f"LLaMA: {example['ru_llama3_1_8b']}\n")

Sources

PEACH – Healthcare corpus (72.8%)
MeSpEn_Glossaries – Medical glossaries (16.6%)
Misraj/Tarjama-25 – General translation corpus (9.7%)
NAMAA-Space/ASCAT – Scientific abstracts (1.0%)

Citation

If you use this dataset, please cite:

@misc{arabic_russian_sci_corpus,
  author = {ArabicNLPWorld Community},
  title = {Arabic–Russian Scientific Translation Corpus (Gemma & LLaMA)},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ArabicNLPWorld/arabic-russian-scientific-translations}
}

License

This dataset is released under the MIT License.

Downloads last month: 17

Total file size:

20.7 MB