ramachandrajoshi/english-kannada-cleaned

English–Kannada Cleaned

A cleaned parallel corpus of English–Kannada sentence pairs suitable for training and evaluating machine translation models.

• Languages: English -> Kannada
• License: Apache License 2.0

Dataset statistics

• Train: 8,00,000 sentence pairs
• Validation: 1,000 sentence pairs
• Test: 1,000 sentence pairs
• Total: 5,02,000 sentence pairs

These counts exclude per-file CSV headers.

Source and provenance

The dataset is provided as UTF-8 CSV files with two columns: english_sentences and kannada_sentences. The data appears cleaned for common noisy artifacts and includes sentence-aligned pairs.

Directory layout:

• train/ — 30 CSV files (train_part_1.csv ... train_part_50.csv) each with header english_sentences,kannada_sentences.
• validation/val.csv — validation split with header.
• test/test.csv — test split with header.

Example row from test/test.csv:

english_sentences	kannada_sentences
No one understood what was going on.	ಏನು ನಡೆಯುತ್ತಿದೆ ಎಂಬುದು ಯಾರಿಗೂ ಅರ್ಥವಾಗಲಿಲ್ಲ.

Recommended usage

You can load the dataset locally using the datasets library (it will read the CSV files directly):

from datasets import load_dataset

data_files = {
        "train": "train/*.csv",
        "validation": "validation/val.csv",
        "test": "test/test.csv",
}

dataset = load_dataset("csv", data_files=data_files)

# access columns
print(dataset["train"][0])

License

This dataset is released under the Apache License 2.0. See the LICENSE file for details.

Citation

If you use this dataset, please cite it. A CITATION.cff is included with suggested metadata.

---

Acknowledgements

• Thanks to the original CSV provider damerajee/en-kannada for sharing the parallel English–Kannada data used as a source for this cleaned dataset.
• Thanks to NSP and AI4Bharat for supporting the creation of this dataset and for providing access to current best open-weight English→Kannada translation LLM models.

---