Shared Task & Competition

Welcome to “MMLoSo Language Challenge 2025”! India is land of unmatched diversity, especially in terms of spoken languages. Many of these languages are tribal languages, which have a sizable number of speakers. But often, these languages are poorly documented and thus lack the massive annotated corpora that power today’s NLP breakthroughs. Limited digital presence means the native speakers face barriers in healthcare messaging, disaster alerts, e-governance, and educational resources—all of which increasingly rely on text mining and machine translation.

By building open-source systems for LRL ⇆ HRL translation, this competition channels deep‐learning skills toward tangible social impact: making vital information accessible in underserved languages and amplifying the voices of native speakers online. In this competition you will translate between high‐resource languages (HRL) and our focused low‐resource languages (LRL), i.e.:

Hindi ⇆ Bhili
Hindi ⇆ Mundari
Hindi ⇆ Gondi
English ⇆ Santali

By tackling all tracks together, you will help push the frontier of multilingual NLP while building end‐to‐end pipelines that work in data‐sparse settings.

Dataset at a Glance

File	Purpose	Key Columns
`bhili-train.csv`	Data for Bhili – Hindi translation training	`row_id`, `hindi`, `bhili`
`gondari-train.csv`	Data for Gondi – Hindi translation training	`row_id`, `hindi`, `gondi`
`mundari-train.csv`	Data for Mundari – Hindi translation training	`row_id`, `hindi`, `mundari`
`santali-train.csv`	Data for Santali – English translation training	`row_id`, `english`, `santali`
`test.csv`	Unlabeled source sentences to translate (released later)	`row_id`, `source_sentence`, `source_lang`, `target_lang`

All texts are drawn from a private, permissively licensed source, cleaned and curated for research.
Courtesy: Ministry of Tribal Affairs, Government of India.

Tasks & Expected Outputs

Task	What you submit	Where	Metric
Machine Translation	`submission.csv` with columns: `row_id`, `source_lang`, `source_sentence`, `target_lang`, `target_sentence`	On Kaggle	BLEU & chrF (tokenized, case-insensitive)

Evaluation Metric

Leaderboard ranks teams by a weighted composite score:

```text Final Score = 0.6 × BLEU + 0.4 × chrF

Why 0.6 / 0.4? Translation quality is harder to push in low-resource settings; the higher weight reflects its research importance.

Rules & Timeline

Standard Kaggle Code Competition rules apply.
External data allowed if publicly available before 30th Aug 2025 and linked in your write-up.
Trainin data release: 16th August 2025
Training Phase: 17th Aug - 5 October 2025
testing Phase: 5th Ocotober - 15th October 2025
Team merger deadline: TBD
Final submission deadline: TBD
Private leaderboard reveal: TBD

See the Rules tab on the competition page for full details.
Happy modeling – and thank you for advancing NLP for underrepresented languages!

📂 Data

1. `bhili-train.csv`

Sentences for supervised Bhili – Hindi translation.

Column	Type	Description
`row_id`	int	Unique row identifier
`hindi`	str	Sentence in the high-resource language (Hindi)
`bhili`	str	Gold translation in the low-resource language

2. `mundari-train.csv`

Sentences for supervised Mundari – Hindi translation.

Column	Type	Description
`row_id`	int	Unique row identifier
`hindi`	str	Sentence in the high-resource language (Hindi)
`mundari`	str	Gold translation in the low-resource language

3. `gondari-train.csv`

Sentences for supervised Gondi – Hindi translation.

Column	Type	Description
`row_id`	int	Unique row identifier
`hindi`	str	Sentence in the high-resource language (Hindi)
`gondi`	str	Gold translation in the low-resource language

4. `santali-train.csv`

Sentences for supervised Santali – English translation.

Column	Type	Description
`row_id`	int	Unique row identifier
`english`	str	Sentence in the high-resource language (English)
`santali`	str	Gold translation in the low-resource language

5. `test.csv`

Unlabeled source sentences. Participants must predict the target_sentence column.

Data Provenance, Licensing, & Copyrights

Machine Translation parallel corpora distilled from publicly released web-crawls and Wikipedia dumps, post-processed using the NGO-Aligned filtering toolkit.
All text is redistributed under Creative Commons BY-SA 4.0. Use outside this competition must cite the original sources.

Citation

If you publish work using this dataset, please cite:

```bibtex @misc{lrlchallenge2025, title = {Multimodal Models for Low-Resource Contexts and Social Impact 2025}, year = {2025}, howpublished = {Kaggle Competition}, url = {https://www.kaggle.com/competitions/mmloso2025} }