This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence (CC BY-NC-SA 4.0): research, academic, personal and other non-commercial use is permitted with attribution; adaptations must be shared under the same licence; commercial use is not permitted. The original public source data remains under the Government Open Data License - India (GODL-India). Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

BharatData-Audit

Licence: CC BY-NC-SA 4.0 — research and non-commercial use only; derivatives must be shared under the same licence. Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai

Access is approved automatically when the email you enter matches your Hugging Face account email and belongs to the organisation you name; other requests are reviewed manually.

An arithmetic and consistency audit of 266,707 tables on data.gov.in. Each finding says where a table does not add up (components vs totals, male + female vs total, percentages out of range, duplicate keys, ten-fold year-on-year jumps, two tables that differ on the same cell, or a data.gov.in value that differs from NITI Aayog's NDAP for the same place, year and indicator).

findings: 272,768 (severity info / warn / flag); tables with at least one flag: 5477
OGD vs NDAP: 342 tables matched to 146 NDAP datasets by value; 1552 place-years differ
scorecards.*: per-ministry share of identities that hold, digit heaping and Benford second-digit tests

Each finding points to the exact table, row and column to review.

Quickstart

The repository is gated: accept the terms on the dataset page and log in with huggingface-cli login first.

# pip install pandas pyarrow duckdb huggingface_hub
import duckdb
import pandas as pd
from huggingface_hub import snapshot_download

root = snapshot_download("judicialmind/BharatData-Audit", repo_type="dataset")

findings = duckdb.sql(f"SELECT * FROM read_json_auto('{root}/findings.jsonl')")
print(duckdb.sql("SELECT severity, count(*) AS n FROM findings GROUP BY 1 ORDER BY 2 DESC").df().to_string(index=False))

# the largest sum-to-total flags, with the table they point at
top = duckdb.sql("""SELECT uuid, ministry, rel_diff, message FROM findings
                    WHERE "check" = 'sum_to_total' AND severity = 'flag' ORDER BY observed DESC LIMIT 1""").df()
print(top.iloc[0].to_dict())

score = pd.read_parquet(f"{root}/scorecards.parquet")
print(score.nlargest(3, "tables")[["ministry", "tables", "identity_hold_rate", "share_tables_warn_or_flag"]].to_string(index=False))

Output:

Findings by severity

Severity	Findings
info	215,285
warn	31,956
flag	25,527

Largest sum-to-total flag

Field	Value
Resource id	8cb79743-4a94-418e-869d-47367e976778
Ministry	Rajya Sabha
Relative difference	0.4214
Message	column 'Funds Approved': the 'Grand Total' row is 30,470,234,559 but its components add up to 52,657,851,980; does not add up

Top 3 publishers by number of tables

Ministry	Tables	Identity hold rate	Share of tables with warn or flag
Ministry of Health and Family Welfare	139,355	0.9963	0.0304
Rajya Sabha	28,869	0.9726	0.0425
Ministry of Jal Shakti	28,196	0.9943	0.0009

The source table of a finding is https://api.data.gov.in/resource/<uuid> (or https://ndap.niti.gov.in/dataset/<id> for ndap-<id>); rows are 0-based row indices of that table and cols its column names.

evaluate.py (pandas + pyarrow) reproduces the numbers on this card and scores other detectors:

python evaluate.py summary                 # headline numbers, recomputed from findings/tables/cross-source files
python evaluate.py recall                  # detector recall on synthetic injected errors
python evaluate.py table 8cb79743-4a94-418e-869d-47367e976778   # all findings for one table
python evaluate.py compare my_findings.jsonl   # your detector vs BharatData-Audit warn/flag findings

compare reads JSON lines with uuid, severity and optionally rows / cols, and prints table- and cell-level agreement (share of your detections that BharatData-Audit also raised) and coverage (share of BharatData-Audit's warn/flag findings you raise).

Files

file	content
`findings.jsonl`	272,768 findings (schema below)
`tables.parquet`	the 266,707 audited tables: `uuid, title, ministry, source` (`ogd` 260,101 / `ndap` 6,589 / `HMIS state reports (data.gov.in)` 17), `template_id`, `n_rows, n_cols`, `error`, `evaluated_by_check` and `held_by_check` (JSON: identities tested and held per check), `n_findings`
`scorecards.parquet`, `scorecards.csv`, `scorecards.md`	per ministry (227): tables, identities tested / held and hold rate, tables with a warn or flag, findings per check, digit heaping (share of last digits 0 or 5), Benford second-digit MAD and conformity, NDAP-aligned columns and agreement
`digits.json`	last-digit and second-digit counts per ministry (values >= 50)
`crosssource_aligned.jsonl`	2,243 data.gov.in columns aligned by value with an NDAP series: `uuid, ndap_id, ogd_col, ndap_series, scale` (unit factor), `n_common, n_agree, share, ministry`
`tagger_tags.jsonl`	30,000 model-suggested readings of tables whose structure the rules could not resolve: `total_rows`, `total_cols`, `components` (total column -> component columns), `notes`, `ok`, `raw`; advisory only, never a finding by itself
`audit_summary.json`, `crosssource_summary.json`, `summary.json`	run totals: findings by check, identities evaluated / held per check, cross-source pairing counts
`injection_recall.json`	synthetic error-injection recall per error kind and per check
`evaluate.py`	summary, recall, per-table inspection and detector comparison

`findings.jsonl` schema

field	meaning
`id`	finding id
`uuid`	table id: data.gov.in resource id, `ndap-<id>`, or `hmis-state-...`
`source`	`ogd`, `ndap`, `HMIS state reports (data.gov.in)`
`check`	which check fired (see below)
`severity`	`info` (relative difference <= 1% or absolute <= 2, or a descriptive test), `warn` (<= 5%), `flag` (> 5%) for identity checks
`rows`, `cols`	0-based rows and column names the finding points at
`expected`, `observed`, `rel_diff`	the value the identity implies, the published value, their relative difference
`message`	one-sentence explanation
`ministry`, `template_id`	publisher of the table and its template (tables sharing a layout)
`other_uuid`	the other table for cross-table checks, `ndap-<id>` for `ogd_vs_ndap`
`extra`	check-specific details (e.g. the other table's row, the year)

Checks: sum_to_total (components vs a Total row or column), mf_total (male + female vs total), ru_total (rural + urban vs total), of_which ("of which" component larger than its parent), pct_range (percentage outside 0-100), pct_count (percentage vs count / base), impossible_value (negative values in an otherwise non-negative column), duplicate_rows, duplicate_key (same key, different values), yoy_jump (ten-fold jump between years, within or across tables), template_conflict (two copies of one table that differ on a cell), ogd_vs_ndap (data.gov.in vs NDAP for the same place, year and indicator), digit_heaping and benford2 (digit tests, info).

Results

Findings by check (python evaluate.py summary):

check	info	warn	flag
sum_to_total	59,983	24,834	17,738
of_which	3,642	2,175	6,917
mf_total	179	27	65
ru_total	5	2	5
pct_range	-	187	209
pct_count	292	88	111
impossible_value	-	3,465	-
duplicate_rows	63,636	-	-
duplicate_key	24,565	-	-
yoy_jump	-	629	178
template_conflict	4	1	7
ogd_vs_ndap	707	548	297
digit_heaping	8,547	-	-
benford2	53,725	-	-
all	215,285	31,956	25,527

Tables with any finding: 70,191; with a within-table flag: 5,477 (5,619 counting cross-table and NDAP flags).

Coverage

266,707 tables: 260,101 data.gov.in resources, 6,589 NDAP datasets and 17 further government health tables, from 227 publishers (ministries, departments, State governments); snapshot May 2026. Tables with personal fields are not audited. Identities were tested in 266,083 tables. Cross-source pairs: 25,366 candidate data.gov.in / NDAP pairs, 2,243 columns aligned.

Source data

Derived from 266,707 resources published on the Open Government Data Platform India (data.gov.in), snapshot May 2026, under the Government Open Data License - India. Each table's publisher (ministry), title and resource id (uuid) are in tables.parquet. GODL-India does not cover personal information; tables with personal fields were excluded.

Providers: Agriculture and Farmer's welfare, Govt of Meghalaya, Andhra Pradesh, Animal Husbandry and Veterinary Department, Govt of Meghalaya, Assam, Backward class welfare, Govt of Karnataka, Central Electricity Authority, Commerce & Industries Department, Govt of Meghalaya, Comptroller And Auditor General of India(CAG), Comptroller and Auditor General of India, DBT Mission, Cabinet Secretariat, Delhi (NCT), Department for Promotion of industry and Internal Trade, Department for Promotion of industry and Internal Trade, Ministry of Commerce and Industry, Department for Promotion of industry and Internal Trade,Ministry of Commerce and Industry,Office of the Economic Adviser, Department of Agriculture and Farmers Welfare, Department of Agriculture, Govt of Karnataka, Department of Animal Husbandry and Dairying, Department of Animal Husbandry, Govt of Karnataka, Department of Atomic Energy, Department of Cooperatives, Govt of Karnataka, and 207 more.

Cross-source comparisons also use the NITI Aayog National Data and Analytics Platform (ndap.niti.gov.in).

Original data: published by the 227 ministries, departments and State governments named per table in tables.parquet (largest: Ministry of Health and Family Welfare, Rajya Sabha, Ministry of Jal Shakti, Ministry of Home Affairs, Ministry of Education) on the Open Government Data Platform India (data.gov.in) and, where marked, NITI Aayog's National Data and Analytics Platform (NDAP), under the Government Open Data License - India (GODL-India). The original public data remains under GODL-India; the CC BY-NC-SA 4.0 licence of this repository covers JudicialMind's compilation, derived fields, annotations, benchmarks and models.

Citation

@misc{judicialmind2026ogdaudit,
  title={BharatData-Audit: An arithmetic and consistency audit of India's open government data tables},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/BharatData-Audit}}
}

License

CC BY-NC-SA 4.0 (see LICENSE): research, academic, personal and other non-commercial use, with attribution to JudicialMind and to the original data publishers; adaptations must be shared under the same licence. Commercial use requires a separate licence.

Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.