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.

HMIS-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.

Cell-level data-quality findings for India's Health Management Information System (HMIS) sub-district monthly reports, April 2020 to March 2022, and a reporting-consistency score for each of 704 districts. Every finding marks a reported value as implausible: a part larger than its whole, public + private not adding up to rural + urban, the same value copied forward for months, a level far from the series median, digit heaping, a value far outside a Chronos-2 forecast band, or a high score from a LightGBM detector trained on injected errors. The data cannot tell a reporting error from fabrication, so findings say "implausible", never "false". The detector model ships with the dataset (detector_lgbm.txt) together with detector.py to score new cells.

Quickstart

import pandas as pd

# after `huggingface-cli download judicialmind/HMIS-Audit --repo-type dataset --local-dir HMIS-Audit`
rules = pd.read_parquet("HMIS-Audit/findings/rules.parquet")   # or "hf://datasets/judicialmind/HMIS-Audit/findings/rules.parquet"
print(rules.groupby(["check", "severity"]).size().unstack().to_string())

trust = pd.read_parquet("HMIS-Audit/district_trust.parquet")
names = pd.read_parquet("HMIS-Bench/series/all.parquet", columns=["district_lgd", "district_name", "state_name"]
                        ).drop_duplicates("district_lgd")      # district names from HMIS-Bench
low = trust.merge(names, on="district_lgd").sort_values("trust_score").head(3)
print(low[["district_name", "state_name", "n_cells", "flag_rate", "trust_score"]].round(4).to_string(index=False))

Example output:

Findings by check and severity

check	flag	warn
copy_forward	8,027	13,571
level_shift	344,399	148,124
of_which	3,274	11,197
ru_total	105	6

Three lowest-trust districts

district	state	cells	flag rate	trust score
Mumbai Suburban	Maharashtra	1,648	0.0768	0.0
Firozabad	Uttar Pradesh	26,832	0.0536	0.1
Jamui	Bihar	37,773	0.0501	0.3

Score cells with the detector

detector_lgbm.txt is a LightGBM booster (binary classifier, 12 features). detector.py builds the features from a long monthly table and applies the booster and its threshold (threshold_fpr1 in detector_eval.json, the score that flags 1% of clean cells):

pip install lightgbm==4.7.0 pandas==3.0.6 numpy==2.3.5 pyarrow==26.0.0     # or: pip install -r requirements.txt

import sys
sys.path.insert(0, "HMIS-Audit")
from detector import score, load_detector          # load_detector() -> lightgbm.Booster

cells = pd.read_parquet("HMIS-Bench/series/all.parquet",
                        columns=["series_id", "ts", "target", "district_lgd", "indicator"])
res = score(cells)                                 # adds score (0-1), flag (bool), severity
print(len(res), "cells scored,", int(res.flag.sum()), "above the 1%-FPR threshold,",
      int((res.severity == "flag").sum()), "with score > 0.9")
print(res.sort_values("score", ascending=False).head(3)[["series_id", "ts", "target", "score"]]
      .round(3).to_string(index=False))

5840115 cells scored, 60069 above the 1%-FPR threshold, 841 with score > 0.9
                     series_id         ts  target  score
  hmis|18|295|2051|9.7.2|total 2022-02-01   105.0    1.0
hmis|10|190|1345|4.1.1.b|total 2020-09-01   115.0    1.0
  hmis|27|500|4092|9.1.6|total 2022-01-01    95.0    1.0

This reproduces findings/detector.parquet (60,069 cells, 841 of them flag). Input: one row per (series, month) with series_id, ts, target (count), district_lgd and indicator; a series is one sub-district × indicator. From the shell: python detector.py cells.parquet --out scores.parquet.

The 12 features, computed per series: f_log_v (log10 value), f_lr_med, f_lr_prev, f_lr_next (log ratio to the series median, the previous and the next month), f_eq_prev, f_eq_next (same value as the neighbour month, for values ≥ 10), f_last0, f_last5 (value ≥ 50 ending in 0 or 5), f_share_dev (log ratio of the cell's share of its district total to its median share), f_robust_z (distance from the median in MAD units), f_mad_rel (MAD / median) and f_month.

Files

file	rows	content
`findings/rules.parquet`	528,703	identity, indicator-chain, copied-forward and level-shift findings
`findings/digits.parquet`	2,182	digit heaping and Benford second-digit findings per sub-district and district
`findings/exceed.parquet`	6,808	values outside the Chronos-2 forecast band, August 2021 to January 2022
`findings/detector.parquet`	60,069	cells above the detector threshold
`district_trust.parquet`	704	district reporting-consistency score
`digits.parquet`	5,720	last-digit counts and digit-test statistics per sub-district (5,031) and district (689)
`detector_lgbm.txt`		the LightGBM detector
`detector_eval.json`		detector evaluation on injected errors and its threshold
`detector.py`, `requirements.txt`		feature code and scoring

Schema

All four findings/*.parquet files share one schema, one row per finding:

column	meaning
`id`	stable finding id
`check`	`ru_total`, `of_which`, `copy_forward`, `level_shift`, `digit_heaping`, `benford2`, `quantile_exceedance`, `detector`
`severity`	`flag` (strong), `warn`, `info`
`indicator`	HMIS indicator code (2017 format); chain name for `of_which`; `subdistrict` / `district` for digit checks
`expected`, `observed`	expected and reported value (meaning per check, below)
`rel_diff`	relative difference; for `detector` the detector score
`message`	plain-language description
`x_state_lgd`, `x_district_lgd`, `x_subdistrict_lgd`	LGD codes of the cell
`x_ts`	month (null for digit checks, which cover all months)
`x_rule`	the rule or parameter: chain name, `run_len=<n>`, `median_ratio`, digit level, forecast model, `lgbm`

Checks:

ru_total: public + private differs from rural + urban (flag above 5% of the total).
of_which: a part is larger than its whole beyond a tolerance; chains (child ≤ tolerance × parent): first-trimester ANC ≤ ANC registrations, C-sections ≤ institutional deliveries, low birth weight ≤ newborns weighed, newborns weighed ≤ 1.05 × live births, live births ≤ 1.05 × deliveries, immunisation sessions held ≤ planned, positive malaria smears ≤ smears tested (monthly); Penta3 ≤ 1.2 × Penta1, fully immunised ≤ 1.5 × BCG, deliveries ≤ 1.5 × ANC registrations (per financial year). Parts of at least 10 are checked; flag when the part exceeds twice the limit. expected = tolerance × parent, observed = child.
copy_forward: the same value ≥ 10 reported 3 or more months in a row (flag from 4).
level_shift: a value at least 5× above or below the series median (value or median ≥ 50; flag from 10×).
digit_heaping: counts ≥ 50 end in 0 or 5 in more than 35% of cases (20% expected) with a last-digit chi-square p < 0.001; benford2: second-digit mean absolute deviation from Benford above 0.012.
quantile_exceedance: the reported value lies more than one band width outside the 0.1-0.9 band of the zero-shot Chronos-2 one-month-ahead forecast on HMIS-Bench (value or median ≥ 10; flag further out, beyond twice the half-band on that side).
detector: detector score above threshold_fpr1 (flag above 0.9).

district_trust.parquet: state_lgd, district_lgd, n_cells (reported cells, all 320 indicators), n_subdistricts, weighted_flags, flag_rate (= weighted flags / cells), by_check (finding counts per check), trust_score, rank_in_state (1 = fewest flags), n_in_state. Trust score = 100 × (1 − percentile rank of the flag rate among all districts): 100 = fewest implausible reports, 0 = most. Each finding is weighted by severity (flag 1, warn 0.5, info 0.1) and by check (digit_heaping 2, benford2 0.5, all others 1).

digits.parquet: level, state_lgd, district_lgd, subdistrict_lgd, n (counts ≥ 50 tested), heaping_share, last_digit_chi2_p, benford2_mad, last_0 ... last_9.

Results

Detector evaluation on injected errors (detector_eval.json): synthetic errors were injected into about 3% of the cells of a sample of real series; the detector was trained on three error kinds and tested on held-out series with all five kinds, including two kinds it never saw. Threshold = 1% false-positive rate on clean test cells (score 0.1776).

injected error	trained on	test cells	AUROC	AUPRC	recall at 1% FPR
spike (value × 3-10)	yes	1,404	0.9936	0.7288	0.8846
drop (value × 0.05-0.3)	yes	1,007	0.9916	0.4870	0.8054
copy forward (previous month's value)	yes	1,030	0.9930	0.2747	0.7388
heap (rounded to 50 or 100)	no	762	0.5910	0.0036	0.0157
digit swap (last two digits)	no	959	0.6872	0.0141	0.0886
all kinds		264,203 test cells	0.8768	0.5280
unseen kinds (heap + digit swap)			0.6446	0.0147

The detector finds large spikes, drops and copied values; rounding and swapped digits look like normal values cell by cell and are covered by the digit tests instead.

Findings by check:

check	flag	warn	info
`level_shift`	344,399	148,124
`copy_forward`	8,027	13,571
`of_which`	3,274	11,197
`ru_total`	105	6
`quantile_exceedance`	5,477	1,331
`detector`	841	59,228
`digit_heaping`		4
`benford2`		693	1,485

of_which findings by chain: fully immunised ≤ BCG 4,185; sessions held ≤ planned 3,832; live births ≤ deliveries 3,288; newborns weighed ≤ live births 1,456; deliveries ≤ ANC 769; first-trimester ANC ≤ ANC 582; Penta3 ≤ Penta1 178; low birth weight ≤ weighed 100; positive smears ≤ tested 58; C-sections ≤ institutional deliveries 23.

Coverage

HMIS sub-district monthly reports, April 2020 to March 2022 (24 months): 5,622 sub-districts in 704 districts of 33 States/UTs, 320 indicators for the rule checks and digit tests, the 48 HMIS-Bench headline indicators for the detector (5,840,115 cells), the 100,000 HMIS-Bench series for the forecast-band check.

Source data

Derived from the HMIS sub-district monthly reports (indicator-wise, FY2020-21 and FY2021-22) of the Ministry of Health and Family Welfare, published on the Open Government Data Platform India (data.gov.in), snapshot May 2026, under the Government Open Data License - India. GODL-India does not cover personal information; the source holds facility-reported aggregate counts only.

Providers: Ministry of Health and Family Welfare (HMIS).

Original data: published by Ministry of Health and Family Welfare (Health Management Information System, HMIS) 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{judicialmind2026hmisaudit,
  title={HMIS-Audit: Data-Quality Findings and District Trust Scores for Indian HMIS Reports},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/HMIS-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.

Downloads last month: 25

Total file size:

11.7 MB

Collection including judicialmind/HMIS-Audit

India Open Government Data — Bharat Health Datasets

Collection

Research and non-commercial use (CC BY-NC-SA 4.0). Commercial: surya@judicialmind.ai / hello@judicialmind.ai • 2 items • Updated 4 days ago