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.

BharatTab

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.

A benchmark of 17 district-level tabular regression tasks for India, built from BharatDistrictPanel: crop yields, HMIS health indicators, PM-KISAN, UDISE+ school statistics and Udyam MSME registrations. Each task predicts one district indicator from hundreds of other administrative indicators of the same district (the target's own source family is excluded). Spatial tasks hold out the districts of about 20% of states; temporal tasks hold out the latest year. Every task ships train/test tables, the tuned LightGBM baseline model with its feature importances, and evaluate.py, which reproduces the leaderboard from these files and scores new models.

Headline result: BharatTab-FM (tabicl_ft) beats the tuned LightGBM baseline on 15 of the 17 tasks.

Quickstart

import lightgbm as lgb
from datasets import load_dataset
from huggingface_hub import hf_hub_download

task = "udyam_registrations"
ds = load_dataset("judicialmind/BharatTab", task)            # gated: accept the terms, then `hf auth login`
train, test = ds["train"].to_pandas(), ds["test"].to_pandas()
feats = [c for c in train.columns if c.startswith("f:") or c.startswith("lag")]
print(train.shape, test.shape, len(feats))

booster = lgb.Booster(model_file=hf_hub_download("judicialmind/BharatTab", f"models/{task}/lightgbm.txt",
                                                 repo_type="dataset"))
pred = booster.predict(test[feats].to_numpy(float))          # one prediction per test row, on the scale of `y`
y = test["y"].to_numpy()
print(f"{task}: LightGBM test R2 = {1 - ((y - pred) ** 2).sum() / ((y - y.mean()) ** 2).sum():.3f}")

Output: train (627, 404), test (135, 404), 400 features; udyam_registrations: LightGBM test R2 = 0.943.

Evaluation harness (CLI):

hf download judicialmind/BharatTab --repo-type dataset --local-dir BharatTab
pip install -r BharatTab/requirements.txt
python BharatTab/evaluate.py                                   # mean, ridge, lightgbm, lag1 on all 17 tasks
python BharatTab/evaluate.py --predictions my_preds.parquet --name my_model    # score your model

evaluate.py prints one line per task and model (R2, the leaderboard R2, RMSE, MAE, n) and the largest R2 difference to leaderboard.json; with the shipped files it reproduces every mean, ridge, lightgbm and lag1 entry (max |R2 - leaderboard R2| = 4.26e-12). --models catboost,tabicl and --models tabicl_ft --tabicl-ckpt bharattab_fm.ckpt (BharatTab-FM) re-fit the other leaderboard models (needs catboost / tabicl; on CPU they match the leaderboard to 3 decimals).

Submitting a model. Write one row per test row of each task: a parquet or csv file with columns task, district_lgd, year, pred (pred on the scale of y, i.e. log1p for tasks marked log1p below). Score it with evaluate.py --predictions, then open a discussion on this repository with the file, the printed table and a short model description to be added to the leaderboard. Fit only on train.parquet of each task.

Leaderboard (test R2)

task	kind	mean	ridge	lightgbm	catboost	tabicl_ft	lag1
crop_yield_maize_kharif	temporal	-0.048	0.424	0.760	0.743	0.821	0.798
crop_yield_mustard_rabi	temporal	-0.001	0.508	0.548	0.515	0.607	0.340
crop_yield_rice_kharif	temporal	-0.009	0.678	0.735	0.704	0.777	0.705
crop_yield_wheat_rabi	temporal	-0.010	0.630	0.616	0.454	0.801	0.883
hmis_c_section_share	spatial	-0.158	0.448	0.867	0.829	0.811
hmis_institutional_delivery_share	spatial	-0.055	-0.659	0.786	0.797	0.814
hmis_penta3_penta1_ratio	spatial	-0.235	-0.929	-0.781	-1.411	-0.182
hmis_sex_ratio_at_birth	spatial	-0.009	-0.010	-0.785	-0.104	-0.021
pmkisan_beneficiaries_next_year	temporal	-0.017	0.634	0.742	0.710	0.745	0.734
pmkisan_female_share	spatial	-2.957	-4.702	-0.805	-0.801	-0.816
udise_enrolment_next_year	temporal	-0.001	-2.055	0.948	0.950	0.981	0.993
udise_girls_toilet_share	spatial	-2.061	-0.899	-1.878	-1.343	0.092
udise_govt_enrolment_share	spatial	-0.198	-0.193	0.200	0.235	0.389
udise_ptr_next_year	temporal	-0.007	0.191	0.885	0.902	0.908	0.915
udise_pupil_teacher_ratio	spatial	-0.334	-0.444	0.619	0.685	0.685
udise_school_electricity_share	spatial	-0.101	-0.986	0.017	0.079	0.018
udyam_registrations	spatial	-0.109	-0.836	0.943	0.940	0.971

mean = training-target mean; ridge = standardised ridge regression; lightgbm = the tuned models in models/; catboost = CatBoost with default-style settings; tabicl_ft = BharatTab-FM; lag1 = the target's value one year earlier (temporal tasks only). RMSE and MAE for every entry, and further reference models, are in leaderboard.json.

Tasks

task	kind	target	log1p	target year(s) train / test	train rows	test rows	features
`crop_yield_maize_kharif`	temporal	`crop:('Maize', 'kharif')`	no	2016-2018 / 2019	1,047	201	720
`crop_yield_mustard_rabi`	temporal	`crop:('Rapeseed &Mustard', 'rabi')`	no	2015-2017 / 2018	1,067	363	716
`crop_yield_rice_kharif`	temporal	`crop:('Rice', 'kharif')`	no	2015-2017 / 2018	947	317	710
`crop_yield_wheat_rabi`	temporal	`crop:('Wheat', 'rabi')`	no	2015-2017 / 2018	1,083	359	714
`hmis_c_section_share`	spatial	`hmis.c_section_share`	no	2021	581	130	400
`hmis_institutional_delivery_share`	spatial	`hmis.institutional_delivery_share`	no	2021	583	129	400
`hmis_penta3_penta1_ratio`	spatial	`hmis.penta3_penta1_ratio`	no	2021	587	130	400
`hmis_sex_ratio_at_birth`	spatial	`hmis.sex_ratio_at_birth`	no	2021	586	130	400
`pmkisan_beneficiaries_next_year`	temporal	`pmkisan.beneficiaries`	yes	2022-2024 / 2025	2,214	759	524
`pmkisan_female_share`	spatial	`pmkisan.female_share`	no	2025	623	136	400
`udise_enrolment_next_year`	temporal	`udise.enrolment`	yes	2016-2018 / 2019	2,005	701	743
`udise_girls_toilet_share`	spatial	`udise.share_schools_with_functional_girl_toilet`	no	2019	574	124	400
`udise_govt_enrolment_share`	spatial	`udise.govt_enrolment_share`	no	2019	577	124	400
`udise_ptr_next_year`	temporal	`udise.pupil_teacher_ratio`	no	2017-2018 / 2019	1,348	701	606
`udise_pupil_teacher_ratio`	spatial	`udise.pupil_teacher_ratio`	no	2019	577	124	400
`udise_school_electricity_share`	spatial	`udise.share_schools_with_functional_electricity`	no	2019	574	124	400
`udyam_registrations`	spatial	`udyam.registrations`	yes	2026	627	135	400

Spatial tasks: one cross-section at the target's latest well-covered year; test rows are all districts of a fixed ~20% of states, so train and test share no state. Temporal tasks: train rows are earlier target years, test rows the latest year; features are taken from the year before the target year and include the target's own lag1 / lag2. Targets marked log1p are stored and scored as log1p(value).

Files

path	content
`tasks/<task>/train.parquet`, `tasks/<task>/test.parquet`	the task tables (schema below)
`tasks/<task>/task.json`	task definition: `name`, `kind`, `target`, `log1p`, `excluded_families` (feature families left out), `n_train`, `n_test`, `n_features`, `years`
`models/<task>/lightgbm.txt`	tuned LightGBM baseline (LightGBM text format; `lgb.Booster(model_file=...)`), inputs = the feature columns in table order
`models/<task>/importances.csv`	`feature`, `gain`: split-gain importance of every feature of the LightGBM model
`leaderboard.md`, `leaderboard.json`	test results per task and model (`r2`, `rmse`, `mae`, `n`, `n_train`, `n_features`)
`evaluate.py`	evaluation harness (reproduce the baselines, score submissions)
`requirements.txt`	pinned dependencies of `evaluate.py`

Schema

column	type	description
`district_lgd`	int32	district code (Local Government Directory)
`state_lgd`	int16	state code (Local Government Directory)
`year`	int16	year of the target value
`y`	float64	target (log1p-transformed when the task is marked log1p)
`f:<indicator>`	float64	one feature per BharatDistrictPanel indicator: its latest value up to the feature year for that district; NaN when not reported. Families: `ndap.<dataset>.<indicator>` (NITI Aayog NDAP), `udise.`, `hmis.`, `mandi.`, `eshram.`, `pmkisan.`, `udyam.`
`lag1`, `lag2`	float64	target value 1 and 2 years before `year` (temporal tasks only)

Coverage

Indian districts: 357 to 768 districts from 27 to 35 states and union territories per task (crop tasks 27, HMIS 33, UDISE+, PM-KISAN and Udyam 35); 17 tasks, 20,287 rows in total, 400 to 743 features per task. Target years: crop yields 2015-2019, HMIS 2021, UDISE+ 2016-2019, PM-KISAN 2022-2025, Udyam 2026.

Source data

Source data: Open Government Data Platform India (data.gov.in) and NDAP (ndap.niti.gov.in), used under the Government Open Data License - India (GODL-India), https://data.gov.in/government-open-data-license-india. Aggregates only; no personal data.

Original data: published by the Ministry of Health and Family Welfare (HMIS), Ministry of Education (UDISE+), Ministry of Agriculture and Farmers Welfare (PM-KISAN, crop statistics), Ministry of Micro, Small and Medium Enterprises (Udyam) and the other ministries and departments whose district indicators appear as features (via BharatDistrictPanel) 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{judicialmind2026bharattab,
  title        = {BharatTab: District-Level Tabular Prediction Tasks from Indian Open Government Data},
  author       = {JudicialMind},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/judicialmind/BharatTab}}
}

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: 23

Total file size:

36.7 MB

Collection including judicialmind/BharatTab

India Open Government Data — Bharat Forecasting Benchmarks

Collection

Research and non-commercial use (CC BY-NC-SA 4.0). Commercial: surya@judicialmind.ai / hello@judicialmind.ai • 1 item • Updated 3 days ago