---
license: other
language:
- en
- hi
task_categories:
- document-question-answering
- text-retrieval
tags:
- parliament
- india
- ocr
- public-policy
pretty_name: Sansad Corpus
---

# Sansad Corpus

A provenance-preserving corpus of Lok Sabha and Rajya Sabha documents for
non-commercial research. Publications are uploaded in bounded tranches and do
not imply completeness unless a release explicitly says so.

The sources are the [Digital Sansad](https://sansad.in/) question APIs and the
[Parliament eLibrary](https://elibrary.sansad.in/). Each document is mapped to
an official source record and its original PDF bytes by SHA-256. The archive
retains additional eLibrary ORIGINAL-bundle PDFs, including Hindi variants
when offered; those attachments are linked in the manifest but are not yet
separately text-extracted.

Each tranche provides selected official PDFs, page-level Markdown, plain text
and JSON in WebDataset TAR shards, Zstandard-compressed JSONL, Parquet tables,
a DuckDB snapshot, and SHA-256 checksums. `manifest.jsonl` identifies the
source records, bitstreams and files. The bundle verifier checks every included
original PDF against its recorded SHA-256; publication also verifies the
uploaded file set and content IDs.

## Text layers and coverage

The canonical layer currently defaults to LiteParse native/OCR text. A separate
free vision-model transcription is retained per page when available; it does
not silently replace the local layer. Page JSON records route, engine,
validation flags and model provenance. Machine-generated text can contain
errors, especially in scans and tables, and should be checked against the PDF
for consequential use.

A tranche may be a pilot, partial session, or complete snapshot. Historical
eLibrary completion refers to a dated census snapshot, not a guarantee that
the live source cannot change. Older tranches may predate attachment archiving;
use each tranche's metadata and manifest rather than assuming uniform coverage.

## Provenance and limitations

Original Parliament material retains its source rights and must be
appropriately attributed. This dataset does not assert that all original
material is public domain. The texts are research derivatives, not official
transcripts; treat validation flags and review provenance as part of the data.