BERTaav — Hindi Political Discourse Corpus

Dataset Summary

BERTaav is a Hindi-language corpus for domain-adaptive pre-training (DAPT) and research on Hindi political discourse and argument mining. It follows the corpus-building approach of RooseBERT (Dore et al., 2026) for English political debates. It has two configurations that differ in source and register, and they should be described and used separately:

Config	Source	Documents	Text	Register
default	Lok Sabha (lower house of the Indian Parliament) Questions and Answers, OCR'd from PDFs	85,775 (77,198 train / 8,577 validation)	~0.89 GB cleaned text	Formal parliamentary Q&A, 1950s–2025
sangraha_political	Politically-relevant subset of the Hindi part of Sangraha (AI4Bharat), selected with a keyword-based filter	2,912,463 (2,883,035 train / 29,428 validation)	~15 GB	Web text, mostly political news and commentary

To our knowledge, the default config is the first large-scale Hindi parliamentary corpus released for NLP research. The sangraha_political config is much larger, is not parliamentary text, and is selected by a heuristic filter whose precision is only informally estimated (see Known limitations under that config below). If you combine the configs for pre-training, consider down-weighting sangraha_political so that the parliamentary data is not swamped.

Possible uses include DAPT of Hindi language models, argument mining (claim/premise detection), stance detection, named-entity recognition for Indian political entities, and policy classification.

Loading the data

from datasets import load_dataset

# Lok Sabha Q&A (default config); keep only good OCR
lok = load_dataset("Shubh1198/BERTaav", split="train")
lok_ok = lok.filter(lambda x: x["quality_flag"] == "ok")

# Sangraha political subset (~5.6 GB download) -- streaming avoids downloading everything
sang = load_dataset("Shubh1198/BERTaav", "sangraha_political", split="train", streaming=True)
core = sang.filter(lambda x: x["tier"] == "core")   # stricter ~56% of the config

---

Config default: Lok Sabha Questions and Answers

Splits

A 90/10 train/validation split, stratified by source type, as in RooseBERT. Documents flagged as low quality are kept in both splits so nothing is silently discarded; filter on quality_flag (for example == "ok") to exclude them.

Split	Documents
train	77,198
validation	8,577

Data fields

Field	Type	Description
id	string	Unique document identifier (DSpace item UUID)
source_type	string	Source category (lok_sabha_qa)
source_collection	string	Collection name on eLibrary Sansad
title	string	Question title, parsed from the document header
date	string	Date of the parliamentary session (YYYY-MM-DD)
question_type	string	unstarred, starred, or unknown
question_number	int	Official Lok Sabha question number
questioner	string	Name of the MP who raised the question
ministry_parsed	string	Responding ministry, parsed from the header
language	string	Always hi (Hindi)
source_pdf	string	Original PDF filename from eLibrary Sansad
source_pdf_bytes	int	Size of the source PDF in bytes
ocr_engine	string	OCR engine used (e.g. tesseract-5.2.0)
ocr_confidence_mean	float	Mean Tesseract confidence (0–100) across all pages
num_pages	int	Number of pages in the source PDF
text	string	Full OCR-extracted Hindi text of the document
text_bytes_utf8	int	Size of the extracted text in UTF-8 bytes
text_hash	string	MD5 hash of the text (for deduplication)
quality_flag	string	ok, low_confidence, low_text, or duplicate
quality_notes	string	Semicolon-separated list of quality issues found
pages_json	string	JSON array of per-page text and confidence scores
processed_at	string	ISO timestamp of post-processing

Example record

{
  "id": "46abc1d6-15f4-47f9-826a-b9225193e75d",
  "source_type": "lok_sabha_qa",
  "source_collection": "Part 1 (Questions And Answers)",
  "title": "मोल्दोवा में विशेष दूतावास कार्यालय",
  "date": "2022-03-25",
  "question_type": "unstarred",
  "question_number": 3843,
  "questioner": "श्री टी.एन. प्रथापन",
  "ministry_parsed": "विदेश मंत्रालय",
  "language": "hi",
  "source_pdf": "AU3813_hindi.pdf",
  "source_pdf_bytes": 34719,
  "ocr_engine": "tesseract-5.2.0",
  "ocr_confidence_mean": 92.74,
  "num_pages": 1,
  "text": "भारत सरकार\nविदेश मंत्रालय\n\nलोक सभा\n...",
  "text_bytes_utf8": 4371,
  "quality_flag": "ok"
}

Source and coverage

All documents come from the eLibrary Sansad archive maintained by the Lok Sabha Secretariat (elibrary.sansad.in), collection Part 1 (Questions And Answers), which holds 1,155,268 items. Only Hindi PDFs were used (filename suffix _hindi.pdf or dc.language.iso=hi); 85,786 of them entered the OCR pipeline, roughly 7% of the collection. Coverage runs from the First Lok Sabha (1950s) to the Eighteenth (2025), across all ministries, for both starred and unstarred questions.

Collection methodology

1. Crawling. The DSpace 7 REST API (/server/api/discover/search/objects) was paginated to enumerate the collection; Hindi PDF bitstreams were identified by filename pattern and downloaded.
2. PDF to image. PDFs were converted to 300 DPI PNG images with pdf2image (poppler) on an HPC cluster using 50 parallel PBS array jobs.
3. OCR. Tesseract 5.2.0 with the Hindi language pack, LSTM engine (--oem 3) and automatic page segmentation (--psm 3). Per-page confidence scores were stored alongside the text.
4. Post-processing. Each document was cleaned and assigned a quality flag. Documents affected by each cleaning step:

Cleaning step	Documents affected
Trailing border-noise artifacts stripped	52,044
Lone-punctuation artifacts stripped	6,295
Garbled URLs replaced with [URL]	2,473
Whitespace normalised	1,004
मूल हिंदी में OCR header artifact removed	602

Quality statistics

Counts for the 85,775 released documents (3 documents with empty text were excluded):

Quality flag	Count	%
ok (mean confidence ≥ 70)	82,613	96.3%
low_confidence (mean confidence < 70, typically scanned tables)	3,063	3.6%
duplicate (identical text hash to an earlier document)	87	0.1%
low_text (< 100 bytes extracted)	12	<0.1%

Mean OCR confidence across all 85,778 processed documents:

Confidence band	Documents
90–99	30,691
80–89	42,707
70–79	9,313
60–69	1,959
50–59	752
< 50	356

Known OCR artefacts (documented, not corrected):

• Four-digit years occasionally lose their leading 1 (e.g. 1918 → 918) when scan quality degrades slightly.
• Mixed-script content (English acronyms, website URLs) may be partially garbled.
• Tabular content (annexures, statistical tables) produces low-confidence output and is the main source of low_confidence documents.

---

Config sangraha_political: political subset of Sangraha

Source

Sangraha (AI4Bharat; Khan et al., 2024). We used its verified/hin Hindi subset: 17,420,932 documents, about 97 GB. According to the Sangraha documentation, the verified subset combines text scraped from human-verified websites, OCR of Indic-language PDFs, and transcribed speech. Our own spot-checks found mainly news articles and commentary, plus some OCR'd book text.

How the subset was selected

The filter is a transparent keyword heuristic, not a trained classifier:

1. Tokenise the first 20,000 characters of each document into Devanagari word chunks (the danda । is treated as punctuation).
2. Score each document: 100 × (2 × strong_hits + weak_hits) / n_tokens, using whole-word matching against a hand-curated vocabulary of about 135 terms: 80 strong terms (government and office-holders, legislature, elections and parties, constitution and democracy, e.g. मतदान, विपक्ष, भाजपा, कांग्रेस, लोकसभा) and 55 weak terms (policy, administration, courts, accountability, e.g. योजना, आयोग, अदालत). Ambiguous words (for example सुरक्षा, सीमा, आधार, रोजगार) and person or place names were deliberately left out. Words such as उम्मीदवार (candidate) and रैली (rally) are weak terms, because they also appear in exam notices and non-political events.
3. Keep a document if it has at least 80 tokens, at least 2 distinct strong terms, and a score of at least 4.0.
4. Exact deduplication on an MD5 of the whitespace-normalised text (first occurrence kept).
5. Drop documents starting with Posted On, the boilerplate of Press Information Bureau (PIB) releases, which are collected separately for this project.

Stage	Documents
Sangraha verified/hin	17,420,932
After political filter (18.49%, 16.52 GB of text)	3,221,339
After exact deduplication (295,026 duplicates removed, 9.16%)	2,926,313
After dropping PIB-style documents (13,850)	2,912,463

Tiers

Every document has a tier so users can pick their own precision/volume trade-off:

Tier	Rule	Documents
core	≥ 3 distinct strong terms and score ≥ 8	1,642,098
extended	all other kept documents (≥ 2 distinct strong terms, score ≥ 4)	1,270,365

Split	Documents	core	extended
train	2,883,035	1,625,564	1,257,471
validation	29,428	16,534	12,894

The validation split (about 1%) is chosen deterministically from the document hash, so it is reproducible and contains no duplicates of training documents. It is not stratified.

Data fields

Field	Type	Description
text	string	Document text, as provided by Sangraha
source	string	Always sangraha_hindi
n_tokens	int	Devanagari word chunks in the first 20,000 characters (see note below)
distinct_strong	int	Number of distinct strong terms found
political_score	float	The density score defined above
tier	string	core or extended
text_md5	string	MD5 of the whitespace-normalised text

n_tokens and political_score are stored per document, so stricter or looser subsets can be recreated without reprocessing the raw data. Summing n_tokens gives about 1.08 billion word-level tokens (1,064,640,320 train + 10,830,430 validation). This is a lower bound: only the first 20,000 characters of each document are counted, and a language-model tokenizer will give different numbers.

Known limitations (sangraha_political)

• Precision is only informally checked. In a manual read of 30 kept documents at the final threshold, about 22 were clearly political news or commentary. The remainder were borderline: government announcements and press-release text, exam and civics material, and encyclopedic pages. This is not a formal evaluation, and recall was not measured.
• Recall gap for policy commentary. Analytical or legal and economic commentary that uses policy vocabulary but few party-political words (for example critiques of a budget) is often rejected.
• Register. The text is news, reporting, commentary and some textbook material, not debate or Q&A. It differs in style from the parliamentary default config.
• Vocabulary bias. The term list centres on national, electoral politics and the main national parties. Local, regional and non-electoral political discourse is likely under-represented.
• Duplicates. Only exact duplicates were removed. Syndicated near-duplicates remain, and some documents repeat paragraphs internally. PIB-style releases were removed only by the Posted On prefix, so some PIB-derived text may remain.
• Noise. Some documents come from OCR'd books, with older orthography and OCR errors.

---

Intended use

Recommended

• Continued pre-training of Hindi encoder models (e.g. MuRIL, IndicBERT) on political-domain text, with the default config as the core parliamentary data.
• Argument-mining, stance and policy-classification research in Hindi. The question→ministry-answer structure of the default config gives implicit argumentative structure. There is currently no annotated Hindi political argument-mining dataset.
• Ablations on core versus extended data.

Out of scope

• Training generative models to fabricate parliamentary statements or political news.
• Presenting documents as verbatim official records without checking the primary source, since OCR introduces errors.
• Profiling or targeting individuals. sangraha_political contains news that names private individuals (for example crime reports), and no personal-data removal was performed.

Licensing

The compilation, processing pipeline, metadata schema and quality annotations are released under CC BY 4.0.

• default: documents are official records of the Parliament of India published by the Lok Sabha Secretariat on eLibrary Sansad. Users are responsible for confirming the terms of reuse of the source material with the Secretariat.
• sangraha_political: derived from Sangraha, which is distributed by AI4Bharat under CC BY 4.0. The text itself originates from third-party web sources, and its copyright status is not addressed by the Sangraha documentation. Users should assess this for their own use.

Roadmap

Item	Status
Lok Sabha Q&A (85,775 documents)	Released
Sangraha political subset (2.9M documents)	Released
Constituent Assembly Debates (Hindi, 167 PDFs)	OCR pending
PIB press releases (176,064 cleaned documents)	Collected; deduplication in progress
Near-duplicate removal across all sources	Planned
Rajya Sabha Q&A	Planned
Annotated Hindi argument-mining gold standard	Planned

Citation

@dataset{agarwal2026bertaav,
  author    = {Agarwal, Shubham},
  title     = {{BERTaav}: A Hindi Political Discourse Corpus for Domain-Adaptive Pre-Training},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/Shubh1198/BERTaav},
  note      = {Lok Sabha Q\&A from eLibrary Sansad; political subset of AI4Bharat Sangraha}
}

Please also cite the works this corpus builds on:

@article{khan2024indicllmsuite,
  title   = {IndicLLMSuite: A Blueprint for Creating Pre-training and Fine-Tuning Datasets for Indian Languages},
  author  = {Mohammed Safi Ur Rahman Khan and Priyam Mehta and Ananth Sankar and
    Umashankar Kumaravelan and Sumanth Doddapaneni and Suriyaprasaad G and
    Varun Balan G and Sparsh Jain and Anoop Kunchukuttan and Pratyush Kumar and
    Raj Dabre and Mitesh M. Khapra},
  year    = {2024},
  journal = {arXiv preprint arXiv: 2403.06350}
}

@article{dore2026roosebert,
  title   = {{RooseBERT}: A New Deal For Political Language Modelling},
  author  = {Dore, Deborah and Cabrio, Elena and Villata, Serena},
  journal = {arXiv preprint arXiv:2508.03250},
  year    = {2026}
}

Dataset Card Authors

Shubham Agarwal, iPhD, Ashoka University shubham.agarwal_phd24@ashoka.edu.in

Acknowledgements

Processing was carried out on the HPC cluster at Ashoka University. Lok Sabha data is provided by the Lok Sabha Secretariat via eLibrary Sansad; the political subset is derived from AI4Bharat's Sangraha.