ShrutamMT-v1-data

Cleaned, multilingual-tagged training data for ShrutamMT-v1 — an 11 Indic languages → English translation model.

⚠️ Full credit: AI4Bharat

This dataset is entirely derived from AI4Bharat Samanantar, the largest publicly available parallel corpora collection for 11 Indic languages, created and released by the AI4Bharat lab. All data collection, mining, and curation credit belongs to them. This repository only applies:

1. Length filtering (1–400 characters per side)
2. Length-ratio filtering (0.3x–3.0x between source and target)
3. Reformatting into a single combined multilingual X → English direction with source-language tags (e.g. <hi>, <bn>, <ta>)
4. Train/validation split (1000 held-out sentences per language for validation)

No new parallel sentences were collected — this is purely a cleaned/reformatted derivative for training a single joint multilingual model.

Contents

File	Description
train.src	Source side: native-language text, prefixed with a language tag, e.g. <hi> यह एक वाक्य है
train.tgt	Target side: corresponding English translation
val.src / val.tgt	Held-out validation split (1000 pairs per language, 11,000 total)

Languages and pair counts (training split)

Language	Code	Pairs
Assamese	as	136,793
Bengali	bn	8,528,272
Gujarati	gu	3,039,408
Hindi	hi	10,002,540
Kannada	kn	4,062,458
Malayalam	ml	5,867,398
Marathi	mr	3,604,472
Odia	or	991,960
Punjabi	pa	2,950,622
Tamil	ta	5,151,342
Telugu	te	4,900,754
Total		49,236,019

Usage

with open("train.src", encoding="utf-8") as f_src, open("train.tgt", encoding="utf-8") as f_tgt:
    for src_line, tgt_line in zip(f_src, f_tgt):
        # src_line looks like: "<hi> यह एक वाक्य है"
        # tgt_line looks like: "This is a sentence"
        pass

Citation

@article{ramesh2021samanantar,
  title={Samanantar: The Largest Publicly Available Parallel Corpora Collection for 11 Indic Languages},
  author={Ramesh, Gowtham and others},
  journal={Transactions of the Association for Computational Linguistics},
  year={2022}
}