randomDude26/mi_kh_tulu_dataset

Low-Resource Indic Machine Translation: Mizo, Khasi, Tulu, Gondi

Parallel corpora for four low-resource Indian languages, released with decontaminated validation and test splits.

The validation and test sets have been filtered so that no exact or near duplicate of any test or validation sentence appears in the training data.

Languages

Pair	Source	Target	Target script	Family
en-mizo	English	Mizo (lus)	Latin	Tibeto-Burman
en-khasi	English	Khasi (kha)	Latin	Austroasiatic
en-tulu	English	Tulu (tcy)	Kannada	Dravidian
hi-gondi	Hindi	Gondi (gon)	Devanagari	Dravidian

Splits

Pair	Train	Validation	Test	Mean test length (src / tgt, tokens)
en-mizo	49,496	464	800	19.3 / 22.3
en-khasi	22,358	482	800	19.4 / 29.6
en-tulu	16,218	488	800	19.5 / 15.5
hi-gondi	15,500	389	737	15.5 / 13.4

Test sets are standardised to 800 sentences. Gondi retains 737 — after decontamination it has only 737 clean sentences, and reaching 800 would require re-admitting contaminated data.

Format

Tab-separated, no header, no quoting, one pair per line:

source<TAB>target

Files are stored quote-free: no field contains a tab, newline, or CSV quote-wrapping, so they parse identically under any reader. Read them with quoting=csv.QUOTE_NONE (pandas: quoting=3).

import pandas as pd
df = pd.read_csv("MIZO/ENG-MIZO/benchmark.tsv", sep="\t", header=None,
                 names=["source", "target"], quoting=3)

or via datasets:

from datasets import load_dataset
ds = load_dataset("randomDude26/mi_kh_tulu_dataset", "en-tulu")

Decontamination

Starting from the full 1,000-sentence test set per pair, text was normalised (Unicode NFC; unified quotes and dashes; repair of tokenisation artefacts such as 8. 65% → 8.65%, via. com → via.com). A test sentence was then removed if any of the following held:

Criterion	Description
exact_pair_leak	the (source, target) pair appears verbatim in train
exact_src_leak	the source appears verbatim in train
tgt_leak_train	the target appears verbatim in train
near_src_leak	source ≥ 0.90 character-similarity to a train source
ngram_leak	≥ 0.80 of the source's word 4-grams occur in train sources
jaccard_leak	token-Jaccard ≥ 0.80 against some train source
val_*_leak	overlaps the validation set on either side
intra_dup	duplicates another test sentence

Matching is on a normalised key (whitespace-collapsed, lowercased, quotes stripped). The same filter was applied to the validation sets.

Test set removals

Criterion	en-mizo	en-khasi	en-tulu	hi-gondi
exact (source, target) in train	0	1	13	1
exact source in train	107	132	9	103
exact target in train	0	0	6	91
near-duplicate source (≥0.90)	2	1	11	53
4-gram containment ≥0.80	0	0	2	4
token-Jaccard ≥0.80	0	0	0	4
source or target in validation	3	2	0	2
duplicate of another test row	0	1	1	5
total removed	112	137	42	263
clean set	888	863	958	737
released (standardised)	800	800	800	737

Validation set removals

Pair	Original	Removed	Released
en-mizo	500	36	464
en-khasi	500	18	482
en-tulu	499	11	488
hi-gondi	500	111	389

Verification

After filtering, all four splits are mutually disjoint. Every cell below is 0:

	exact pair	exact src	exact tgt	near src ≥0.90	4-gram ≥0.80
test vs train	0	0	0	0	0
test vs val	0	0	0	0	0
val vs train	0	0	0	0	0

Per-row removal logs are in decontamination/, listing every dropped sentence, the criterion that dropped it, and the training sentence it matched — so the filtering is fully auditable and reproducible.

Repository layout

MIZO/ENG-MIZO/     train.tsv  val.tsv  benchmark.tsv   # benchmark.tsv = test
KHASI/ENG-KHASI/   train.tsv  val.tsv  benchmark.tsv
TULU/ENG-TULU/     train.tsv  val.tsv  test.tsv
GONDI/HIN-GON/     train.tsv  val.tsv  test.tsv
Monolingual/       *_mono_cpt_train.tsv  *_mono_perplexity_val.tsv
decontamination/   per-row removal logs

Monolingual/ holds the corpora used for continued pretraining (CPT) in Mizo, Khasi, Hindi and Kannada. It is monolingual, so the parallel-data decontamination above does not apply to it; the guarantees stated in this card cover the four parallel pairs listed under Splits.

License

This release contains data under two licenses:

Mizo, Khasi, Tulu (MIZO/, KHASI/, TULU/): original parallel corpora created by the authors through native-speaker translation. Released under CC BY 4.0. Gondi (GONDI/): curated and decontaminated from the public CGNetSwara Hindi–Gondi corpus (https://cgnetswara.org/hindi-gondi-corpus.html), which is released under CC BY-NC 4.0. As a derivative of NC-licensed material, the Gondi subset remains CC BY-NC 4.0 — non-commercial use only. Attribution: CGNetSwara, https://cgnetswara.org/hindi-gondi-corpus.html.

Users combining subsets for commercial purposes must exclude the Gondi data or obtain separate permission from CGNetSwara.