README.md

metadata

license: cc-by-4.0
task_categories:
  - translation
language:
  - en
  - mr
size_categories:
  - 1M<n<10M
pretty_name: BhashaSetu v2
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/dev.parquet
      - split: test
        path: data/test.parquet

BhashaSetu v2: English–Marathi parallel corpus

BhashaSetu v2 is maintained at ParamTh/BhashaSetu. This copy is identical; please use and cite the main repository.

6,937,554 English–Marathi sentence pairs, every one checked to be a sentence pair, deduplicated, decontaminated against the standard benchmarks, and labelled with its source, licence and a quality score. Everything is redistributable (CC0 / CC-BY-4.0).

What's new in v2 (October 2026)

• Much larger and fully deduplicated: 6.9M unique pairs, against 24,525 unique pairs in v1's 2.8M rows.
• One sentence pair per row, with original punctuation and casing. v1 rows joined several sentences with punctuation removed.
• Proper held-out splits: validation and test are human translations, disjoint from training and decontaminated against FLORES-200 and IN22. v1's test splits overlapped its training data, so scores measured on them should be re-run on this version.
• Redistributable: built only from CC0 and CC-BY-4.0 sources, released under CC-BY-4.0.
• Per-row provenance and quality scores (source, license, quality_tier, labse).

At a glance

Pairs	6,937,554 (train 6,932,554 / validation 2,000 / test 3,000)
Human-translated pairs	219,086
Mined pairs, LaBSE ≥ 0.85	3,200,251
Mined pairs, LaBSE 0.80–0.85	3,518,217
Candidate rows examined	19,290,824
Licence	CC-BY-4.0 overall; per-row licence in license

The validation and test splits are human translations only. A test set drawn from mined data rewards reproducing the miner's errors.

Loading

from datasets import load_dataset
ds = load_dataset("ParamTh/BhashaSetu")

# Highest-precision subset: human translations plus confidently mined pairs
hq = ds["train"].filter(lambda r: r["quality_tier"] in ("human", "high"))

Columns

Column	Meaning
id	Stable 64-bit hash of the normalised pair
english, marathi	The pair. Original casing and punctuation; never lowercased, stemmed or stripped
source	Which upstream collection the row came from (table below)
license	Licence of that upstream collection
translation	human or mined
quality_tier	human; high (mined, LaBSE ≥ 0.85); standard (mined, 0.80–0.85)
labse	LaBSE cosine for mined rows as released by AI4Bharat (empty for human rows)
en_tokens, mr_tokens	Whitespace token counts

Tiers are descriptive rather than calibrated. A LaBSE threshold tuned on one dataset does not transfer reliably to another, so the score is exposed and the choice is left to you.

Sources

All sources are the English–Marathi portions of AI4Bharat's BPCC (Gala et al., 2023). When a pair occurs in several sources, the row is attributed to the first one in this table.

Source	Description	Translation	Rows kept	Licence
bpcc_wiki	BPCC-H-Wiki: professional human translation of Wikipedia sentences	human	50,386	CC-BY-4.0
bpcc_daily	BPCC-H-Daily: professional human translation of everyday sentences	human	3,771	CC-BY-4.0
bpcc_seed	BPCC seed collection (human-translated; includes NLLB-Seed)	human	60,455	CC-BY-4.0
ilci	ILCI: human-translated health and tourism sentences	human	104,474	CC0-1.0
samanantar_filtered	Samanantar (v0.3), LaBSE-filtered by AI4Bharat	mined	1,262,660	CC0-1.0
samanantar_pp	Samanantar++, mined by AI4Bharat for IndicTrans2	mined	2,911,164	CC0-1.0
nllb_filtered	NLLB mined bitext, LaBSE-filtered by AI4Bharat	mined	2,182,847	CC0-1.0
comparable	Sentences mined from comparable news articles	mined	361,797	CC0-1.0

How it was built

1. Normalisation: markup removal, Unicode and Indic normalisation.
2. Script purity: at least half the English side in Latin script and half the Marathi side in Devanagari.
3. Numeric consistency: every number on the English side appears on the Marathi side.
4. Sentence-boundary gate: rows whose two sides contain different numbers of sentences, rows with more than three sentences, extreme length ratios, and long rows with no internal punctuation are removed. This targets rows where the two sides drift apart partway through, which embedding-similarity filters miss.
5. Hindi check: Hindi text shares the Devanagari script and passes script checks. Rows whose "Marathi" side is dominated by Hindi function words are removed.
6. Benchmark decontamination against FLORES-200 dev and devtest and IN22-Gen: exact matches on either side, plus MinHash near-duplicates of the English side.
7. Deduplication: exact pairs, then at most one row per normalised English sentence and one per normalised Marathi sentence, keeping human translation over mined and then the higher LaBSE score. This removed 965,721 duplicate pairs, 3,286,415 repeated English sentences and 7,317,365 repeated Marathi sentences.

Rows removed at each stage, per source (19,290,824 examined, 18,507,055 passed the gates before deduplication):

Source	Read	Script	Numeric	Boundary	Hindi	Benchmark	Passed
bpcc_wiki	54,348	0	658	1,938	44	6	51,702
bpcc_daily	4,565	0	10	779	4	0	3,772
bpcc_seed	117,275	1	1,160	3,918	92	37	112,067
ilci	116,979	28	3,008	4,876	183	1	108,883
samanantar_filtered	1,366,103	1,085	16,100	57,290	487	4	1,291,136
samanantar_pp	10,806,011	4,044	6,502	200,339	5,120	1	10,590,003
nllb_filtered	6,393,165	23,199	232,871	181,435	3,508	3	5,952,142
comparable	432,378	4,173	14,783	13,811	2,260	0	397,350

One upstream detail worth knowing: BPCC's nllb_filtered score file stores 1 − cosine, unlike its other subsets (our own LaBSE cosines correlate with it at exactly −1.00). We invert it, so labse means the same thing on every row.

Does it train better models?

Fine-tuning NLLB-200-distilled-600M (LoRA, 400 steps, 20,000 rows, seed 42) and evaluating on FLORES-200 devtest (400 sentences):

Training data	BLEU	chrF++	spBLEU
BhashaSetu v1	9.83	42.12	23.08
BhashaSetu v2	12.38	43.70	24.24

This is a small-budget comparison (a 4 GB GPU) that shows the direction of the difference, not its size at scale.

Limitations

• Mined pairs are machine-aligned. The gates remove structural errors, not every mistranslation. Spot checks find wrong pairs mainly among short news headlines in the standard tier, e.g. "Delay in sowing of wheat" paired with a Marathi headline about delayed sugarcane harvesting. For high-precision training use the human and high tiers.
• The sentence segmenter is heuristic and rejects about 4% of clean gold pairs.
• IN22-Conv was not available to us for decontamination.
• The Hindi check uses function words, so short or code-mixed rows can slip through.

Licence and attribution

Released under CC-BY-4.0. The mined and older seed collections were released by AI4Bharat under CC0 and the BPCC-H human translations under CC-BY-4.0; per-row licences are in the license column. Please cite BPCC/IndicTrans2 alongside this dataset:

@article{gala2023indictrans2,
  title   = {IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Languages of India},
  author  = {Gala, Jay and others},
  journal = {Transactions on Machine Learning Research},
  year    = {2023}
}