Health-StatsQA / README.md
sakasurya's picture
Access note: verified organisation email
1090669 verified
|
Raw History Blame Contribute Delete
9.23 kB
---
license: cc-by-nc-sa-4.0
tags:
- health
- india
- question-answering
- hindi
- statistics
- benchmark
- hmis
pretty_name: Health-StatsQA
language:
- en
- hi
task_categories:
- question-answering
- table-question-answering
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: test
path: data/test.jsonl
extra_gated_prompt: '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.'
extra_gated_fields:
Name: text
Email: text
Organization: text
I accept the CC BY-NC-SA licence terms (research and non-commercial use only): checkbox
extra_gated_button_content: Accept and request access
---
# Health-StatsQA
<!-- jm:top -->
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.
<!-- /jm:top -->
1,000 health-statistics questions about Indian districts, each asked in **English and Hindi**
(2,000 questions in all). Every answer is read or computed by code from two sibling releases:
[IndiaDiseaseWatch](https://huggingface.co/datasets/judicialmind/IndiaDiseaseWatch) (HMIS disease counts and
outbreak labels) and the [Bharat Health Service Atlas](https://huggingface.co/datasets/judicialmind/Bharat-Health-Service-Atlas)
(HMIS service indicators, Rural Health Statistics facility counts). Each question comes with the data rows its
answer is computed from, so a model can be tested from memory (`closed_book`) and reading the data (`oracle`).
## Quickstart
```python
from datasets import load_dataset
ds = load_dataset("judicialmind/Health-StatsQA", split="test") # or load_dataset("json", data_files="Health-StatsQA/data/test.jsonl")
q = ds[0]
print(q["lang"], q["template"], q["answer_type"], "->", q["answer"])
print(q["question"])
print(q["oracle_context"])
```
```
en disease_month number -> 2.0
According to India's Health Management Information System (HMIS), how many emergency attendances for snake bite were reported by health facilities in Imphal West district, Manipur in July 2020? Reply with a single number.
emergency attendances for snake bite | Imphal West, Manipur
2020-04: 0
2020-05: 1
2020-06: 2
2020-07: 2
2020-10: 1
2020-11: 0
2020-12: 0
2021-01: 1
2021-02: 0
2021-03: 0
```
### Evaluate a model
```bash
cd Health-StatsQA
python evaluate.py prompts --condition closed_book --out prompts.jsonl # or --condition oracle
# run your model on each prompt, write preds.jsonl lines {"id": ..., "response": "<raw answer text>"}
python evaluate.py score preds.jsonl
python evaluate.py score --gold # parser check: gold answers score 1.0
```
```
all {"n": 2000, "accuracy": 1.0}
answer_type=choice {"n": 400, "accuracy": 1.0}
...
lang=hi {"n": 1000, "accuracy": 1.0}
template=change_pct {"n": 200, "accuracy": 1.0}
...
pairs {"n": 1000, "both_correct": 1.0, "en_hi_agreement": 1.0}
```
`evaluate.py` needs only the Python standard library. The `oracle` prompt is the `oracle_context` rows followed by
the question. Scores: accuracy overall, by language, template and answer type; `pairs.both_correct` = share of
questions answered correctly in both English and Hindi, `pairs.en_hi_agreement` = share where the two languages
are both right or both wrong.
Scoring:
- `number`: the first number in the response (Indian digit grouping, Devanagari digits and a trailing % accepted;
the question's year is skipped when another number is present) is correct within `rel_tolerance` of the answer
(1% for percentages, exact for counts), or when it equals the answer after rounding to the answer's decimals.
- `choice`: the response names the right district and no other choice (case-insensitive).
- `yes_no`: the response answers with the right word (yes / no, हाँ / नहीं).
## Files
- `data/test.jsonl`: 2,000 questions, one per line.
- `sources.parquet`: per-resource attribution (provider, title, URL).
- `evaluate.py`: prompt builder and scorer.
## Schema
| field | meaning |
|---|---|
| `id` | question id (`<pair_id>_<lang>`) |
| `pair_id` | links the English and Hindi versions of a question |
| `question` | question text; district and state names stay in Latin script, as in the LGD directory |
| `lang` | `en` or `hi` |
| `answer` | gold answer: a number, a district name, or yes/no (`yes`/`no` in English, `हाँ`/`नहीं` in Hindi) |
| `answer_type` | `number`, `choice` or `yes_no` |
| `choices` | the options for `choice` and `yes_no` questions |
| `unit` | `count`, `%` or empty |
| `scale` | unit multiplier of the answer (1.0) |
| `rel_tolerance` | 0.01 for percentages, else 0 (exact) |
| `year`, `fy`, `month` | the period asked about (`fy` like `FY2019-20`; `month` as `YYYY-MM-01`) |
| `state_lgd`, `district_lgd`, `level`, `place` | LGD codes, `district`, and the place name |
| `source` | release the answer comes from (`IndiaDiseaseWatch` or `Bharat-Health-Service-Atlas`) |
| `template` | question template (below) |
| `conditions` | `["closed_book", "oracle"]` |
| `split` | `test` |
| `oracle_context` | the data rows the answer is computed from |
## Templates (per language)
| template | questions |
|---|---|
| `change_pct` | 100 |
| `compare_districts` | 100 |
| `disease_fy_total` | 150 |
| `disease_month` | 200 |
| `outbreak_yes_no` | 100 |
| `rhs_count` | 100 |
| `service_ratio` | 150 |
| `state_argmax` | 100 |
How each template's answer is computed:
- `disease_month`: the reported count for one district, disease and month.
- `disease_fy_total`: the financial-year sum, using only years with all 12 months reported.
- `service_ratio`: a ratio from the atlas, in percent.
- `compare_districts`: which of two districts reported more, in the same state and year; the two values differ by
more than 5%.
- `state_argmax`: the district with the most cases among a state's fully-reporting districts, with a lead of more
than 5%.
- `change_pct`: the year-on-year change in percent, from a base of at least 20.
- `rhs_count`: a facility count.
- `outbreak_yes_no`: the IndiaDiseaseWatch label, with the rule stated in the question; balanced between yes and no.
The Hindi questions are written from templates, not machine-translated.
## Coverage
Districts across India; HMIS months from April 2009 to March 2022 (disease and service questions) and Rural
Health Statistics years 2016 to 2021 (facility questions); 2,000 questions in English and Hindi over 8 templates.
## Source data
Derived from 2 resources published on the Open Government Data Platform India (data.gov.in), snapshot May 2026, under the [Government Open Data License - India](https://data.gov.in/government-open-data-license-india).
Per-resource attribution (provider, title, resource URL) is in the `attribution` column of `sources.parquet`.
Providers: Ministry of Health and Family Welfare (HMIS), Ministry of Health and Family Welfare (RHS, via NITI Aayog NDAP).
- Answers derive from IndiaDiseaseWatch and Bharat-Health-Service-Atlas; their `sources.parquet` files list every underlying resource.
<!-- jm:source -->
Original data: published by Ministry of Health and Family Welfare (Health Management Information System, HMIS; Rural Health Statistics, via NITI Aayog NDAP) on the Open Government Data Platform India ([data.gov.in](https://data.gov.in)) and, where marked, NITI Aayog's National Data and Analytics Platform ([NDAP](https://ndap.niti.gov.in)), under the [Government Open Data License - India](https://data.gov.in/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.
<!-- /jm:source -->
## Citation
```bibtex
@misc{judicialmind2026healthstatsqa,
title={Health-StatsQA: Bilingual English-Hindi Health Statistics Questions about Indian Districts},
author={JudicialMind},
year={2026},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/datasets/judicialmind/Health-StatsQA}}
}
```
<!-- jm:license -->
## License
[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/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.
<!-- /jm:license -->