micahchoo/pib-parallel

PIB Parallel

English sentences beside their official translations into 16 Indian languages, from the press releases India's Press Information Bureau (PIB) published, one month to a split: 451,815 sentence pairs and 24,618 document pairs from 4 months so far, each pair traceable to the two releases it came from.

PIB is the Government of India's press office. It publishes each release in English and links the same release translated by its offices. Earlier PIB corpora (CVIT-PIB, PMIndia) end around 2019; this one takes whole months, from every office that publishes in its own language, with release IDs. The months are chosen to differ in time, subject and how much of the translation is machine-made. The pipeline that made it, and fetches any other month, is github.com/micahchoo/pib-parallel.

Before you use it as a test set, read How much is machine translation: part of PIB's translation is a near copy of Google Translate.

Load it

from datasets import load_dataset

sentences = load_dataset("micahchoo/pib-parallel", "sentences", split="nov2025")
documents = load_dataset("micahchoo/pib-parallel", "documents", split="nov2025")
every_month = load_dataset("micahchoo/pib-parallel", "sentences")  # one split per month
tamil = sentences.filter(lambda r: r["lang"] == "ta" and r["similarity"] >= 0.85)

A release is at https://www.pib.gov.in/PressReleasePage.aspx?PRID=<prid>.

What is in it

sentences: one row per distinct sentence pair.

Field	Meaning
lang	the translation's language (ISO 639)
en, text	the English sentence and its translation
similarity	cosine similarity of the two: LaBSE's, 0.70 or more; for Manipuri, LASER3's, 0.80 or more, on a higher scale that is not comparable with LaBSE's
numbers_agree	whether both sides hold the same numbers (lakh, crore, million and billion read as values); null when neither has any
confidence	low for Konkani, Khasi and Manipuri, which LaBSE does not know or knows only through a related language
google_copy	for a sample of pairs only: whether PIB's text is within chrF 90 of Google Translate's for the same English; null where not checked
occurrences	how often PIB published this exact pair ("Friends," opens hundreds of speeches)
en_prid, text_prid	the English release and the translated release
office	the crawl pass that found the pair: delhi-en (English releases and every translation they link) or a regional office (its releases and their English)
date, posted_by	the release's date and the office PIB names as its publisher

documents: one row per English release and one translation, with both titles and bodies (en_title, en_body, title, body), both PRIDs, ministry, office, date, posted_by. Use these to align with another aligner, or to work at document level.

Months

Split	Month	Sentence pairs	Document pairs	English releases
jul2019	July 2019	27,478	2,592	1,103
may2021	May 2021	93,256	6,573	1,009
aug2023	August 2023	144,566	7,157	1,620
nov2025	November 2025	186,515	8,296	1,786

PIB publishes nothing in Bodo, Dogri, Kashmiri, Maithili, Sanskrit, Santali or Sindhi. The offices that publish only their own Hindi (Lucknow, Patna, Bhopal and others) are not crawled.

November 2025

Language	Code	Sentence pairs	Similarity ≥ 0.85	Document pairs
Hindi	hi	34,372	21,809	1,428
Urdu	ur	31,840	22,106	1,354
Gujarati	gu	16,773	12,035	561
Marathi	mr	13,712	7,186	714
Malayalam	ml	13,313	9,483	530
Kannada	kn	12,881	9,373	434
Punjabi	pa	11,165	7,742	457
Telugu	te	9,962	4,649	445
Bengali	bn	9,586	4,256	518
Odia	or	9,522	5,916	306
Assamese	as	8,166	2,824	369
Tamil	ta	7,145	3,704	785
Manipuri (low confidence)	mni	4,702	2,380	228
Konkani (low confidence)	gom	1,467	458	71
Nepali	ne	1,351	966	32
Khasi (low confidence)	kha	558	207	64
All		186,515	115,094	8,296

August 2023

Language	Code	Sentence pairs	Similarity ≥ 0.85	Document pairs
Urdu	ur	28,202	18,892	1,355
Hindi	hi	21,120	13,205	991
Telugu	te	14,067	8,414	717
Marathi	mr	10,844	5,773	688
Tamil	ta	10,518	6,374	812
Kannada	kn	9,657	6,774	363
Punjabi	pa	9,633	6,525	461
Gujarati	gu	8,879	6,261	325
Malayalam	ml	6,974	4,980	285
Odia	or	6,970	4,026	295
Assamese	as	6,059	2,244	241
Manipuri (low confidence)	mni	5,252	2,560	282
Bengali	bn	5,118	1,589	270
Nepali	ne	1,104	630	34
Khasi (low confidence)	kha	169	22	38
All		144,566	88,269	7,157

May 2021

Language	Code	Sentence pairs	Similarity ≥ 0.85	Document pairs
Hindi	hi	19,201	10,984	938
Punjabi	pa	14,490	9,678	758
Urdu	ur	14,171	7,236	888
Telugu	te	8,714	3,652	614
Marathi	mr	7,667	3,238	612
Kannada	kn	5,526	2,919	352
Tamil	ta	4,708	1,354	704
Gujarati	gu	4,293	2,263	250
Bengali	bn	3,732	711	504
Malayalam	ml	3,598	1,589	321
Odia	or	3,253	1,091	237
Manipuri (low confidence)	mni	2,171	838	225
Assamese	as	1,732	299	170
All		93,256	45,852	6,573

July 2019

Language	Code	Sentence pairs	Similarity ≥ 0.85	Document pairs
Urdu	ur	8,682	4,067	763
Hindi	hi	5,334	2,651	434
Marathi	mr	2,675	1,024	262
Bengali	bn	2,570	484	415
Punjabi	pa	1,866	1,028	147
Tamil	ta	1,617	584	175
Gujarati	gu	1,546	687	87
Malayalam	ml	1,074	349	111
Kannada	kn	911	423	78
Telugu	te	825	367	60
Assamese	as	378	57	60
All		27,478	11,721	2,592

Regional

The offices that publish little in their own language give only tens of releases a month, so for them every month PIB has is pooled into one split per language, in the regional and regional_documents configs. Months already published as month splits are left out, so no pair is in both.

nepali = load_dataset("micahchoo/pib-parallel", "regional", split="ne")

Language	Split	Sentence pairs	Similarity ≥ 0.85	Document pairs	Months	Google copies
Konkani (low confidence)	gom	2,909	909	140	November 2023, November 2024	12%
Khasi (low confidence)	kha	8,198	3,186	905	May 2023 to September 2026 (37)	3%
Manipuri (low confidence)	mni	350,496	153,558	25,250	January 2019 to September 2026 (84)	not measured
Nepali	ne	33,364	22,642	1,075	May 2023 to September 2026 (39)	5%

How it was made

1. Fetched from PIB's own pages: every release each office listed for the month, and the translations each release page links. PIB states which releases are translations of which, so no document was matched by guesswork.
2. Cleaned: embedded tweets and posts removed (third-party material, and on translated pages most were left in English; 7,908 in November 2025); email addresses and phone numbers replaced by <email> and <phone> on every side; error pages, withdrawn releases and repeated festival boilerplate dropped.
3. Aligned: each release cut into sentences (not after a title such as "Dr." or "डॉ.", or an initial such as "एल."), embedded with LaBSE, and matched in order, one to one, one to two or two to one, by dynamic programming. Pairs below a similarity of 0.70 were dropped. A pair of documents found from both sides was aligned once.
4. Filtered: a pair was kept only if each side is mostly in its own script. In November 2025 this dropped 22,673 pairs: links, hashtags, staff initials and English lines left on translated pages, which align because they match themselves; number-only table cells; and Hindi speeches quoted on English pages.
5. Checked: every email address and phone number form is searched for again in each month, independently of the rules that removed them, in every script's digits (audit.py). A month with any left is not published.

How much is machine translation

Google Translate translated the English of up to 100 pairs per office and language. A pair whose PIB text is within chrF 90 of Google's counts as a copy. Two independent translations of one sentence almost never come that close: the sarvam-30b model and Google did on 2 of 200 Assamese sentences, and Google and IN22's human translations on 0 of 89 across six languages.

Office, language	November 2018	July 2019	May 2021	August 2023	November 2025
Agartala, Bengali					2%
Ahmedabad, Gujarati	7%	7%	7%	30%	36%
Bengaluru, Kannada	3%	7%	3%	12%	15%
Bhubaneswar, Odia			9%	5%	7%
Chandigarh, Punjabi		3%	8%	3%	17%
Chennai, Tamil	2%	5%	2%	5%	5%
Gangtok, Nepali				0% (of 79)	13%
Guwahati, Assamese	0% (of 57)	0% (of 1)		48% (of 83)	30% (of 64)
Hyderabad, Telugu	0%	0% (of 31)	0% (of 10)	7% (of 68)	0% (of 77)
Kolkata, Bengali	2%	0% (of 22)	0% (of 45)	0% (of 41)	0% (of 88)
Mumbai, Marathi	1%	1%	9%	1%	7%
Shillong, Khasi					12% (of 24)
Thiruvananthapuram, Malayalam	1%		0% (of 1)	4%	13%
Delhi's Assamese translations		0% (of 43)	6%	31%	23%
Delhi's Bengali translations		1%	3%	4%	2%
Delhi's Konkani translations					13%
Delhi's Hindi translations		5%	22%	17%	29%
Delhi's Khasi translations				0% (of 19)	5%
Delhi's Kannada translations			0% (of 4)
Delhi's Malayalam translations		0%	18%	25%	18%
Delhi's Marathi translations		1%	6%	4%	2%
Delhi's Nepali translations				3%	9%
Delhi's Punjabi translations		0% (of 7)
Delhi's Telugu translations		2%	2%	0%	4%
Delhi's Urdu translations		3%	4%	22%	7%

Samples are 100 pairs unless stated; a blank cell was not measured, or the office linked no translations that month. November 2018 is a sample of ten offices, not a published month. The rate is a lower bound on machine text: a heavily edited Google translation, or another engine, passes as "not a copy". PIB's pages are set up to load Bhashini's machine-translation plugin; text made with it would not show as a copy of Google's. Google today also writes differently from Google years ago, so an older month's rate is a lower bound twice over.

What this means for a test set. A score against a Google-made reference favours Google. Use the pairs where google_copy is false, or the offices with the lowest rates (Kolkata, Hyderabad, Chennai, Mumbai), or both. google_copy is known for 6,237 sampled pairs (563 copies), not for every pair: Google's output itself is not in this dataset, as its terms forbid passing it on.

Where PIB links its translations

This method needs PIB to link a release to its translation. Ten sampled releases per office, November of each year:

Office	2017	2018	2019	2020 to 2024
Ahmedabad, Bengaluru, Chennai	10/10	10/10	9-10/10	10/10
Hyderabad	7/10	10/10	10/10	10/10
Guwahati, Kolkata, Mumbai, Thiruvananthapuram	0/10	10/10	10/10	10/10
Chandigarh	0/10	0/10	9/10	10/10
Bhubaneswar	0/10	0/10	0/10	10/10
Delhi (English linking any translation)	0/10	5/10	8/10	10/10
Agartala, Gangtok, Mumbai Konkani, Shillong	no releases until 2023			linked from 2023

From 2020 every office links its translations, so the pipeline covers 2020 onward as built. Before that, the regional pages link their English more often than the English pages link their translations, so crawl from the regional side. Releases that link nothing need retrieval, which the pipeline does not do.

Limits

• A model can have seen it. PIB's text is in many training sets.
• Machine translation in the references, as measured above.
• LaBSE's blind spots. It does not know Manipuri or Khasi and knows Konkani only through Marathi; their pairs are marked low, and Manipuri's are weak (30 of 330 at 0.85 or more in November 2025). For every language, a similarity of 0.70 to 0.85 holds more partial and wrong pairs than above 0.85.
• What the cleaning missed. Tweets pasted into a release as plain paragraphs, rather than embedded, cannot be told from prose (8 of 10,265 pages in November 2025). A Latin-script translation (Khasi) cannot be told from English by its script, so an English line on a Khasi page can remain.
• office is the crawl pass that found the pair, not always the office that translated it: Delhi's English releases link translations that regional offices post. posted_by names the publisher of the release the pass started from.

Terms

The text is PIB's. PIB's Copyright Policy, read on 4 October 2026:

Material featured on this website may be reproduced free of charge and there is no need for any prior approval for using the content. The permission to reproduce this material shall not extend to any third-party material. Authorisation to reproduce such material must be obtained from the departments/copyright holders concerned. The material must be reproduced accurately and not used in a derogatory manner or in a misleading context. Wherever the material is being published or issued to others, the source must be prominently acknowledged.

So: name PIB as the source, and do not describe these translations as human: some are near copies of Google Translate. Embedded third-party posts were removed for this reason.

The annotations (the sentence alignment, similarities, number and script checks, google_copy, occurrences) are licensed CC BY 4.0. The code is MIT, at github.com/micahchoo/pib-parallel.

Cite

@misc{pib_parallel_2026,
  title        = {{PIB Parallel}: English and Indian languages from the Press Information Bureau},
  author       = {Micah},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/micahchoo/pib-parallel}},
  note         = {Text: Press Information Bureau, Government of India. Pipeline: \url{https://github.com/micahchoo/pib-parallel}}
}