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.

IndiaStatsQA

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.

IndiaStatsQA (DistrictBench) has 20,000 numeric questions about Indian state, district and village statistics. Each question is answered by one cell of a government table and is tied to a year and an LGD (Local Government Directory) place code. An answer counts as correct when it is within 10% of the official value. The benchmark measures how much official Indian statistics a language model knows from memory (closed book), and how well it reads the number from the source table (oracle) or from retrieved tables (retrieval).

Quickstart

from datasets import load_dataset

qs = load_dataset("judicialmind/IndiaStatsQA", "questions", split="full")
item = qs[0]
print(item["question"])
print(item["answer"], item["unit"], item["split"])

Output:

According to the Government of India dataset "Item-wise report for Lohardaga of Jharkhand for August-2019-20", what was "SubDistrict - Kisko - Rural [D]" for Lohardaga district, Jharkhand, Indicator: M10 [Number of cases of Childhood Diseases (0-5 years)], Parameters: Childhood Diseases - Diarrhoea, in financial year 2019-20? Reply with a single number.

Answer 59.0, unit count, split test.

With pandas or DuckDB on a local copy:

import pandas as pd, duckdb
qs = pd.read_json("questions_en.jsonl", lines=True)
print(duckdb.sql("SELECT level, split, count(*) AS n FROM qs GROUP BY ALL ORDER BY ALL").df())
scores = pd.read_parquet("items.parquet")          # per-item scores of the leaderboard models
print(scores.groupby(["model", "condition"]).within_10.mean().loc["qwen3-32b"].round(4))

Output:

Questions by level and split

Level	Split	Questions
district	dev	346
district	test	8,343
state	dev	595
state	test	10,620
village	test	96

qwen3-32b, share of answers within 10% by condition

Condition	Within 10%
closed book	0.0208
oracle	0.9314
retrieval	0.2871

The other configs are oracle_contexts and retrieval_contexts (load_dataset("judicialmind/IndiaStatsQA", "oracle_contexts", split="full")).

Evaluate a model

evaluate.py (needs only numpy) builds the prompts, scores answers and rebuilds the leaderboard.

# 1. prompts: {"id", "messages": [system, user]} per line; conditions: closed_book, oracle, retrieval
python evaluate.py prompts --condition oracle --out prompts_oracle_en.jsonl
# 2. run your model on each prompt (greedy decoding), save {"id": ..., "response": "<reply>"} per line
# 3. score
python evaluate.py score --predictions my_model/oracle_en.jsonl --condition oracle --model my-model

score prints one JSON row per subset (all, pre_cutoff, post_cutoff, state, district, village), e.g. {"model": "my-model", "condition": "oracle", "lang": "en", "subset": "all", "n": 16566, "within_10pct": ...}.

The model is expected to reply with one number in the unit the question asks for, or exactly I don't know. The scorer reads the first number in the reply (Indian digit grouping, Indic digits and scale words such as lakh / crore / thousand are handled; a number equal to the question's year is skipped), and treats replies like "I don't know" or "not available" as refusals.

To add a model to the leaderboard, save its predictions as predictions/<model>/<condition>_en.jsonl and run python evaluate.py --out leaderboard.csv; the board is rebuilt from every predictions file.

Reproduce the published leaderboard from the shipped predictions:

python evaluate.py --check leaderboard.csv
# ... one line per model and condition ...
# {"rows": 192, "matches": true, "reference": "leaderboard.csv"}

Files

file	content
`questions_en.jsonl`	20,000 questions with gold answers (schema below)
`oracle.jsonl`	`id`, `context`: the source table excerpt for the item (16,662 items)
`retrieval.jsonl`	`id`, `context`, `hits`: top tables from BM25 search over the full table corpus (resource ids in `hits`)
`predictions/<model>/<condition>_en.jsonl`	raw replies of the 12 leaderboard models: `id`, `response`
`items.parquet`	per-item scores of those replies: `id, model, condition, lang, refused, answered, pred, ale, within_10, score, post_cutoff, level, source, district_lgd, state_lgd, split`
`leaderboard.csv`	aggregate metrics per model, condition, language and subset
`evaluate.py`	prompts, scorer and leaderboard builder

Schema (`questions_en.jsonl`)

field	type	description
`id`	string	item id (`statsqa-item-...`)
`question`	string	the question, naming the source dataset, place, period and unit
`lang`	string	`en`
`answer`	float	official value, in the unit of the question
`unit`	string	unit (`count`, `%`, `persons`, `mm`, `Rs`, ...)
`scale`	float	multiplier from the question's unit to base units (1.0 for most items)
`year`	int	calendar year (start year for financial years)
`fy`	string or null	financial year label such as `2019-20`
`state_lgd`, `district_lgd`	int or null	LGD codes of the place
`level`	string	`state`, `district` or `village`
`source`	string	`ogd` (data.gov.in) or `ndap` (NITI Aayog NDAP)
`source_uuid`	string	data.gov.in resource id, or `ndap-<dataset id>`
`source_cell`	object	`row`, `col` of the answer cell in the source table
`post_cutoff`	bool	source table published on or after 2025-01-01
`conditions`	list	conditions the item is used in (`closed_book`, `oracle`, `retrieval`)
`split`	string	`test` (19,059) or `dev` (941); dev and test never share a source table
`template`	string	question template
`ministry`, `title`	string	publisher and title of the source table
`place`	string	place name as written in the question
`created`	string or null	publication date of the source table

Village items and NDAP items are not asked closed book; oracle contexts exist for items whose answer cell can be shown from a data.gov.in table.

Results

English, all items of each condition (closed book n = 17,904; oracle n = 16,566; retrieval n = 20,000). Prompts are fixed (see evaluate.py), decoding is greedy. within 10%: share of answers within 10% of the official value. MdALE: median |log10(pred/gold)| over answered items (lower is better). refusal score: 1 for a correct answer, 0.25 for "I don't know", 0 otherwise.

model	params (B)	closed book: within 10%	closed book: MdALE	closed book: answer rate	closed book: refusal score	oracle: within 10%	retrieval: within 10%	retrieval: refusal score
qwen3-0.6b	0.75	0.0	0.7501	0.0041	0.2488	0.1305	0.0659	0.2605
qwen3-1.7b	2.0	0.0	0.6968	0.0002	0.2499	0.4331	0.1729	0.3353
qwen3-4b-2507	4.0	0.0115	0.5643	0.1155	0.2326	0.8149	0.2716	0.4434
qwen3-8b	8.2	0.0183	0.7697	0.2928	0.1951	0.7742	0.2658	0.4343
phi-4	14.7	0.0075	0.301	0.0342	0.2489	0.823	0.277	0.4502
qwen3-14b	14.8	0.0158	0.5324	0.1316	0.2328	0.8965	0.2854	0.4515
gpt-oss-20b	20.9	0.0359	0.8004	0.5056	0.159	0.8758	0.2852	0.4581
sarvam-m	23.6	0.0244	0.9379	0.2633	0.2084	0.7983	0.2762	0.4431
mistral-small-3.2	24.0	0.0141	0.2819	0.0729	0.2459	0.8344	0.276	0.4526
qwen3-30b-a3b-2507	30.5	0.042	0.3979	0.3463	0.2054	0.8831	0.2819	0.455
qwen3-32b	32.8	0.0208	0.4683	0.162	0.2303	0.9314	0.2871	0.4568
gpt-oss-120b	116.8	0.0102	0.0858	0.0231	0.2544	0.9328	0.2888	0.4629

Models: Qwen3 (Qwen/Qwen3-0.6B, -1.7B, -8B, -14B, -32B with thinking off; Qwen3-4B-Instruct-2507, Qwen3-30B-A3B-Instruct-2507), microsoft/phi-4, openai/gpt-oss-20b and -120b ("Reasoning: low"), sarvamai/sarvam-m, mistralai/Mistral-Small-3.2-24B-Instruct-2506. Breakdowns by subset (pre/post cutoff, state, district, village) and the disparity ratio (worst-decile / best-decile district error) are in leaderboard.csv and on the IndiaStatsQA-Leaderboard Space.

Coverage

India: 11,215 state-level, 8,689 district-level and 96 village-level items, with places identified by LGD codes; source tables from 64 ministries, departments and State governments; reference years 1948 to 2031 (mostly recent years); 3,468 items come from tables published on or after 2025-01-01.

Source data

Derived from 5,078 resources published on the Open Government Data Platform India (data.gov.in), snapshot May 2026, under the Government Open Data License - India. Per-item source attribution (publisher, table title, resource id) is in the ministry, title and source_uuid fields; the resource page is https://data.gov.in/resource/ followed by the id. GODL-India does not cover personal information; tables with personal fields were excluded.

Providers: Comptroller And Auditor General of India(CAG), Department of Health and Family Welfare, Department of States, Directorate of Tourism, Govt of Meghalaya, Election Commission of India, Forests & Environment Departmnent, Govt of Meghalaya, Haryana, Himachal Pradesh, Ministry of Agriculture and Farmers Welfare, Ministry of Chemicals and Fertilizers, Ministry of Civil Aviation, Ministry of Coal, Ministry of Commerce and Industry, Ministry of Communications, Ministry of Communications,Department of Telecommunications (DOT), Ministry of Consumer Affairs, Food and Public Distribution, Ministry of Cooperation, Ministry of Corporate Affairs, Ministry of Culture, Ministry of Development of North Eastern Region, and 44 more.

2,600 items come from 341 NITI Aayog NDAP datasets (ndap.niti.gov.in).

Original data: published by the ministries, departments and State governments named per item (ministry field) on data.gov.in, and NITI Aayog (NDAP) 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{judicialmind2026indiastatsqa,
  title={IndiaStatsQA: Numeric questions on Indian state and district statistics},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/IndiaStatsQA}}
}

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.