AdiBhashaa

A Community-Curated Benchmark for Machine Translation into Indian Tribal Languages

Pooja Singh¹ · Sandeep Kumar¹

¹Indian Institute of Technology Delhi

---

Dataset summary

Despite substantial advances in multilingual NLP, the languages of India's tribal (Janjatiya/Adivasi) communities are almost entirely absent from public MT benchmarks and foundation models. This digital exclusion exacerbates existing inequities in education, governance, and digital participation.

AdiBhashaa is a community-driven, documentation-oriented Machine Translation benchmark for four under-represented Indian tribal languages. The release contains an 80,000 sentence parallel corpus (Hindi–Tribal) created through a participatory, human-in-the-loop workflow with native speakers.

	Benchmark Details
Total Scale	80,000 parallel sentences (~20,000 per language)
Language Coverage	Bhili, Gondi, Mundari, and Santali
Source Language	Hindi
Domains Covered	Education, Governance, Healthcare, News, Social

Language Representation

	Bhili	Gondi	Mundari	Santali
ISO 639-3	bhb	gon	unv	sat
Family	Indo-Aryan	Dravidian	Austroasiatic (Munda)	Austroasiatic (Munda)
Speakers (2011)	~10.4 M	~3.0 M	~1.1 M	~7.6 M
Script in this release	Devanagari	Devanagari	Devanagari	Ol Chiki

How to load

The datasets library enables you to load and preprocess the curated dataset directly in Python. (Ensure you have an active HuggingFace access token).

from datasets import load_dataset

# Load the entire benchmark
dataset = load_dataset("misniitdelhi/AdiBhasha")

# Alternatively, stream the dataset without downloading
dataset_stream = load_dataset("misniitdelhi/AdiBhasha", streaming=True)
print(next(iter(dataset_stream)))

Dataset structure

Repository layout

data/
├── bhili/
│   └── bhili-train.csv     (~20,565 pairs)
├── gondi/
│   └── gondi-train.csv     (~20,001 pairs)
├── mundari/
│   └── mundari-train.csv   (~20,150 pairs)
├── santali/
│   └── santali-train.csv   (~20,001 pairs)
└── test.csv                (Unified test set)

Dataset creation (Participatory Workflow)

To ensure linguistic quality, cultural fidelity, and community ownership, we adopted a participatory, human-in-the-loop workflow rather than relying solely on crowdsourcing or model-generated text.

1. Source curation: Researchers compile Hindi sentences from public, educational, and civic sources, prioritizing cultural appropriateness and downstream utility.
2. Community translation: Native-speaking translators (experienced teachers or language activists) render the sentences into their languages with an emphasis on semantic adequacy and naturalness.
3. Independent validation: A separate group of validators (early-career researchers from the same communities) reviews each sentence pair for adequacy, fluency, and cultural sensitivity.

Supported tasks & Limitations

• Machine Translation: The primary task is fine-tuning models for Hindi ↔ Tribal language translation.
• Cross-lingual transfer: Studying cross-script transfer from high-resource languages (like Hindi) to closely related writing systems, and modeling for underserved scripts (like Ol Chiki).
• Known Biases: Translating into high-resource languages (Tribal → Hindi/English) systematically outperforms translation into tribal languages due to the difficulty of generating morphologically rich, low-frequency vocabulary. Models trained on this data may exhibit this dominant-language bias.

Licensing

Artefact	Licence
Dataset & Transcripts	CC BY-NC-SA 4.0

Citation

If you use AdiBhashaa in your work, please cite the following paper:

@misc{singh2025adibhashaacommunitycuratedbenchmarkmachine,
      title={AdiBhashaa: A Community-Curated Benchmark for Machine Translation into Indian Tribal Languages}, 
      author={Pooja Singh and Sandeep Kumar},
      year={2025},
      eprint={2512.04765},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.04765}, 
}