This dataset is released for non-commercial, educational and research use only. Access requests are reviewed by the dataset maintainers before being granted.

COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages
**COILD-MT-Corpus
**⚠️ Before requesting access Step 1 (required first): Fill out the COILD dataset access request form. Step 2: Also complete the access-request form on this page (click "Agree and request access" below).
Both steps are required — access requests submitted here without the Google Form on file will not be approved.

The accompanying paper is currently available as a preprint.
Permitted Use: This dataset is intended only for non-commercial, educational and/or research purposes. Users may access, download, process, analyse, and modify the dataset solely for these purposes.
Prohibited Use: Commercial use, sale, sublicensing, or redistribution of the dataset — including modified or derived versions — in whole or in part, is prohibited without prior written authorization from the dataset creators.
Any publication, report, presentation, software, or other research output that uses this dataset must cite: "COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages."

By requesting access, you acknowledge and agree to comply with these terms.

COILD-MT-Corpus is a collection of sentence-aligned parallel text corpora across 20 language pairs, covering Hindi- and Tamil-centric pairings with other Indian languages. Each pair is stored as two line-aligned plain-text files (one sentence per line, same line count on both sides).

Access to this dataset is gated. Requesting access requires agreeing to the non-commercial research-use terms below; requests are reviewed by the maintainers.

Dataset Structure

Each top-level folder is named <SRC>-<TGT> and contains two .txt files, one per language, aligned line-by-line.

Folder	Language pair	Sentence pairs
HIN-ASM	Hindi – Assamese	94,871
HIN-BEN	Hindi – Bengali	73,066
HIN-BRX	Hindi – Bodo	47,599
HIN-DOI	Hindi – Dogri	58,896
HIN-GOM	Hindi – Konkani	49,631
HIN-GUJ	Hindi – Gujarati	96,611
HIN-KAS	Hindi – Kashmiri	58,702
HIN-MAI	Hindi – Maithili	93,748
HIN-MAR	Hindi – Marathi	60,287
HIN-MNI	Hindi – Manipuri (Meitei)	52,494
HIN-NPI	Hindi – Nepali	61,643
HIN-ORY	Hindi – Odia	52,733
HIN-PAN	Hindi – Punjabi	98,163
HIN-SAT	Hindi – Santali	20,603
HIN-SND	Hindi – Sindhi	43,663
HIN-TEL	Hindi – Telugu	64,483
HIN-URD	Hindi – Urdu	84,205
TAM-KAN	Tamil – Kannada	30,273
TAM-MAL	Tamil – Malayalam	23,592
TAM-TEL	Tamil – Telugu	16,033

Total: ~1,181,296 sentence pairs across 20 language pairs.

Usage

The files are plain, newline-delimited, line-aligned text — line N in the source file corresponds to line N in the target file. Example (Python, no extra dependencies):

from huggingface_hub import hf_hub_download

src_path = hf_hub_download(repo_id="coild-dataset/COILD-MT-Corpus", filename="HIN-BEN/Hindi.txt", repo_type="dataset")
tgt_path = hf_hub_download(repo_id="coild-dataset/COILD-MT-Corpus", filename="HIN-BEN/Bengali.txt", repo_type="dataset")

with open(src_path, encoding="utf-8") as f_src, open(tgt_path, encoding="utf-8") as f_tgt:
    pairs = list(zip(f_src, f_tgt))

print(pairs[0])

License and Terms of Use

The COILD dataset is distributed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. The dataset may be used, shared, and adapted exclusively for non-commercial purposes, provided appropriate attribution is given to the dataset creators and the accompanying paper is cited.

Access is additionally gated: requesting access requires agreeing to the terms above and completing the access-request form, including an undertaking to attribute the creators in any resulting work and not to further redistribute the raw dataset without prior permission from the maintainers.

See LICENSE.md for the full terms.

For access requests or related queries, contact iitpainlpmlresourcerequest@gmail.com or deepakkumar1538@gmail.com.

Citation

@inproceedings{COILD-DATASET,
  title     = {COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages},
  author    = {Singh, Kshetrimayum Boynao and
               Mishra, Nitin Kumar and
               Dutta, Palash Pratim and
               Khan, Atai Waris and
               Kaushik, Aparna and
               Kumar, Avinash and
               Deeksha and
               Kumar, Deepak and
               Jha, Saroj Kumar and
               Sengupta, Saloka and
               Bansal, Poonam and
               Dev, Amita and
               Vaidya, Ashwini and
               Singh, Sanasam Ranbir and
               Akhtar, Md Shad and
               Muralikrishna, SN and
               Bhattacharya, Samit and
               Chakraborty, Tanmoy and
               Ekbal, Asif},
  year      = {2026}
}

Maintainers

AI-NLP-ML Research Group (ainlpml-iitp)

Downloads last month: 3

Total file size:

898 MB