opensansad/lok-sabha-qa

OpenSansad — Lok Sabha Q&A Dataset

Structured parliamentary question-and-answer records from India's Lok Sabha

Dataset Summary

This dataset contains 146,300+ parliamentary Q&A records from India's Lok Sabha (lower house of Parliament), covering:

• 18th Lok Sabha — 30,220 questions across sessions 2–7 (Jul 2024 — Mar 2026), fully extracted with question/answer text
• 17th Lok Sabha — 60,548 questions across sessions 1–12, 14–15 (Jun 2019 — Feb 2024), fully extracted with question/answer text
• 16th Lok Sabha — 55,594 questions from sessions 6–17 (Nov 2015 — Feb 2019), fully extracted with question/answer text

Each record contains rich metadata including the responsible ministry, subject, date, MP names, and links to the original source PDFs on Digital Sansad. For records where text extraction is complete, the full question and official government answer are included.

This dataset is part of the OpenSansad initiative by Sammit Jain — a project to make the workings of Sansad (Indian Parliament) more accessible and transparent through open data and open-source tooling.

Supported Tasks

• Question Answering / RAG: The question-answer pairs can be used directly for retrieval-augmented generation over Indian parliamentary proceedings. See the companion lok-sabha-rag project for a working RAG pipeline built on this data.
• Text Classification: Classify questions by ministry, topic, or question type (starred vs. unstarred).
• Summarization: Generate concise summaries of verbose government responses.
• Information Extraction: Extract structured data such as schemes, statistics, and policy details from answer text.
• Multilingual NLP: Some records contain Hindi text, useful for Hindi/English code-mixed NLP tasks.

Languages

The dataset is primarily in English (en), with a subset of records in Hindi (hi). Some records contain a mix of both languages.

Dataset Structure

Data Instance

{
  "id": "LS18-S2-STARRED-280",
  "lok_no": 18,
  "session_no": 2,
  "ques_no": 280,
  "type": "STARRED",
  "date": "2024-08-09",
  "subject": "Role of NGOs in Welfare of Women and Children",
  "ministry": "WOMEN AND CHILD DEVELOPMENT",
  "members": ["Shri Manish Jaiswal"],
  "full_text": "## GOVERNMENT OF INDIA ...",
  "question_text": "*280. SHRI MANISH JAISWAL :\n\nWill the Minister of WOMEN AND CHILD DEVELOPMENT be pleased to state ...",
  "answer_text": "## ANSWER\n\nMINISTER OF WOMEN AND CHILD DEVELOPMENT (SHRIMATI ANNPURNA DEVI)\n\n(a) to (e) : A Statement is laid ...",
  "question_word_count": 147,
  "answer_word_count": 855,
  "pdf_url": "https://sansad.in/getFile/loksabhaquestions/annex/182/AS280_6OmUWJ.pdf?source=pqals",
  "pdf_url_hindi": "https://sansad.in/getFile/loksabhaquestions/qhindi/182/AS280_6OmUWJ.pdf?source=pqals",
  "num_pages": 3
}

Data Fields

Field	Type	Description
id	string	Unique identifier following the pattern LS{lok_no}-S{session}-{type}-{ques_no}
lok_no	int	Lok Sabha number (16, 17, or 18)
session_no	int	Parliamentary session number
ques_no	int	Question serial number within the session
type	string	STARRED (oral answer in Parliament) or UNSTARRED (written answer)
date	string	Date the question was answered, in YYYY-MM-DD format
subject	string	Subject or topic of the question
ministry	string	Government ministry responsible for answering
members	list[string]	Name(s) of the MP(s) who asked the question
full_text	string	Complete text extracted from the source PDF
question_text	string	The question portion, separated from the answer
answer_text	string	The government's answer portion (null for ~0.05% unsplit records)
question_word_count	float	Word count of the question text
answer_word_count	float	Word count of the answer text
pdf_url	string	URL to the original English PDF on Digital Sansad
pdf_url_hindi	string	URL to the Hindi PDF on Digital Sansad
num_pages	float	Number of pages in the source PDF

Data Splits

The dataset is provided as a single train split. Users can create custom splits by lok number, session, date, ministry, or question type as needed.

18th Lok Sabha (2024–2026) — fully extracted

Session	Period	Starred	Unstarred	Total
2	Jul — Aug 2024	280	3,219	3,499
3	Nov — Dec 2024	380	4,369	4,749
4	Feb — Mar 2025	500	5,750	6,250
5	Jul — Aug 2025	419	4,829	5,248
6	Nov — Dec 2025	280	3,219	3,499
7	Jan — Mar 2026	557	6,418	6,975
Total		2,416	27,804	30,220

16th Lok Sabha (2014–2019) — extraction in progress

Session	Period	Starred	Unstarred	Total	Text available
6	Nov — Dec 2015	360	4,135	4,495	4,482 extracted
7	Feb — Mar 2016	280	3,219	3,499	3,497 extracted
8	Apr — May 2016	259	2,990	3,249	3,246 extracted
9	Jul — Aug 2016	399	4,599	4,998	4,991 extracted
10	Nov — Dec 2016	440	5,060	5,500	5,499 extracted
11	Feb — Apr 2017	560	6,439	6,999	6,999 extracted
12	Feb — Mar 2018	377	4,358	4,735	4,735 extracted
13	Mar — Apr 2018	280	3,214	3,494	3,494 extracted
14	Feb — Apr 2018	579	6,668	7,247	7,244 extracted
15	Jul — Aug 2018	360	4,136	4,496	4,491 extracted
16	Dec — Jan 2019	393	4,498	4,891	4,888 extracted
17	Feb — Feb 2019	160	1,831	1,991	1,991 extracted
Total		4,447	51,147	55,594

Note: 16th Lok Sabha extraction is in progress. Remaining sessions will be added incrementally.

17th Lok Sabha (2019–2024) — fully extracted

Session	Period	Starred	Unstarred	Total	Text available
1	Jun — Aug 2019	500	5,698	6,198	6,188 extracted
2	Nov — Dec 2019	380	4,369	4,749	4,746 extracted
3	Jan — Mar 2020	419	4,818	5,237	5,237 extracted
4	Sep — Sep 2020	0	2,301	2,301	2,301 extracted
5	Jan — Mar 2021	440	5,057	5,497	5,493 extracted
6	Jul — Aug 2021	320	3,679	3,999	3,999 extracted
7	Nov — Dec 2021	339	3,909	4,248	4,248 extracted
8	Jan — Apr 2022	500	5,749	6,249	6,249 extracted
9	Jul — Aug 2022	319	3,672	3,991	3,991 extracted
10	Dec — Dec 2022	239	2,759	2,998	2,998 extracted
11	Jan — Apr 2023	480	5,520	6,000	6,000 extracted
12	Jul — Aug 2023	340	3,910	4,250	4,250 extracted
14	Dec — Dec 2023	267	3,066	3,333	3,331 extracted
15	Feb — Feb 2024	120	1,379	1,499	1,499 extracted
Total		4,663	55,886	60,549

Note: Session 13 does not exist in the official Digital Sansad records for the 17th Lok Sabha.

Dataset Creation

Curation Rationale

Indian parliamentary Q&A records are a rich source of information about government policy, public spending, and administrative decisions. However, this data is only available as individual PDF files on the Digital Sansad portal, making it difficult to search, analyze, or use programmatically. This dataset aims to make this information accessible in a structured, machine-readable format.

Source Data

All data is sourced from Digital Sansad (sansad.in), the official portal of the Indian Parliament. The source documents are publicly available PDF files containing parliamentary questions and their official government responses.

Collection process:

1. Question metadata (subject, ministry, members, dates) is scraped from the Digital Sansad search API
2. PDF documents are downloaded from official URLs
3. Text is extracted from PDFs using Docling
4. Question and answer portions are separated using a multi-strategy regex-based splitter (15 strategies covering English, Hindi, and various OCR artifacts)

Personal and Sensitive Information

This dataset contains only publicly available parliamentary records. The names of Members of Parliament and government ministers appear as part of the official record. No private or sensitive personal information is included.

Considerations for Using the Data

Known Limitations

• OCR artifacts: Text is extracted from PDFs, some of which have corrupted embedded text layers. This can result in garbled characters in a small number of records (~0.05% remain unsplit due to severe OCR issues).
• Q/A separation: The question-answer split is automated and may occasionally include header/footer text or split at imprecise boundaries.
• Coverage: Currently covers the 16th (2014–2019), 17th (2019–2024), and 18th (2024–2026) Lok Sabhas. All three are fully extracted (146,300 records with text); the 16th Lok Sabha has sessions 10–17 extracted (39,300 records), remaining sessions in progress.
• Hindi records: A small subset of records are in Hindi. The Q/A separation for Hindi text relies on Hindi-specific markers and may be less reliable than for English records.

Future Improvements

• Complete 16th Lok Sabha extraction: Expand from sessions 10–17 to all sessions
• Earlier Lok Sabhas: Expand coverage to include 15th and older parliamentary data
• OCR fallback: Re-extract garbled PDFs using image-based OCR for the remaining unsplit records
• Minister extraction: Extract and normalize the answering minister's name as a structured field
• Rajya Sabha: Extend to India's upper house of Parliament
• Enrichments: Add MP party affiliation, constituency, and other metadata from external sources

Usage

from datasets import load_dataset

ds = load_dataset("opensansad/lok-sabha-qa")

# Filter by ministry
health = ds["train"].filter(lambda x: "HEALTH" in x["ministry"])

# Filter by session
session_4 = ds["train"].filter(lambda x: x["session_no"] == 4)

# Starred questions only (answered orally in Parliament)
starred = ds["train"].filter(lambda x: x["type"] == "STARRED")

# Search by subject keyword
education = ds["train"].filter(lambda x: "education" in x["subject"].lower())

Use for RAG

This dataset pairs well with retrieval-augmented generation. See the companion lok-sabha-rag project for a complete RAG pipeline that uses this data to answer questions about Indian parliamentary proceedings.

Additional Information

Dataset Curators

This dataset is created and maintained by Sammit Jain as part of the OpenSansad initiative.

Licensing

The dataset is released under CC-BY-4.0. The underlying parliamentary records are public documents of the Government of India.

Citation

@dataset{opensansad_lok_sabha_qa,
  title   = {OpenSansad Lok Sabha Q&A Dataset},
  author  = {Sammit Jain},
  year    = {2026},
  url     = {https://huggingface.co/datasets/opensansad/lok-sabha-qa},
  license = {CC-BY-4.0},
}