You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This model 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-Chronos

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.

HMIS-Chronos is Chronos-2 (120M parameters) fine-tuned to forecast monthly health-service counts that Indian sub-districts report to the Health Management Information System (HMIS): antenatal registrations, deliveries, vaccinations, family planning, disease cases, outpatient and inpatient attendance. It forecasts all indicators of a sub-district together, with Chronos-2 group attention across them, and returns quantiles 1 to 3 months ahead. It is scored on the HMIS-Bench benchmark, where its median forecasts have the lowest MASE of the nine models at every horizon.

Quickstart

pip install -r requirements.txt     # chronos-forecasting 2.3.2, torch, pandas, pyarrow

import pandas as pd
import torch
from chronos import Chronos2Pipeline

pipe = Chronos2Pipeline.from_pretrained("judicialmind/HMIS-Chronos", device_map="cpu", dtype=torch.float32)

# all HMIS-Bench series of one sub-district (Narpatganj, Araria, Bihar), up to July 2021
df = pd.read_parquet("hf://datasets/judicialmind/HMIS-Bench/series/bench.parquet",
                     columns=["series_id", "ts", "target", "subdistrict", "headline"])
df["ts"] = pd.to_datetime(df["ts"])
sub = df[(df.subdistrict == 1112) & (df.ts < "2021-08-01")]
months = pd.period_range(end="2021-07", periods=36, freq="M")       # 36-month context, NaN = not reported
wide = (sub.assign(m=sub.ts.dt.to_period("M"))
           .pivot(index="headline", columns="m", values="target").reindex(columns=months))
inputs = [torch.tensor(row, dtype=torch.float32) for row in wide.to_numpy()]

q, _ = pipe.predict_quantiles(inputs, prediction_length=3, quantile_levels=[0.1, 0.5, 0.9],
                              cross_learning=True)              # group attention across the sub-district
for name, qi in list(zip(wide.index, q))[:2]:
    print(name, "Aug-Oct 2021 median:", qi[0, :, 1].round().tolist(),
          "80% band:", list(zip(qi[0, :, 0].round().tolist(), qi[0, :, 2].round().tolist())))

anc_registered Aug-Oct 2021 median: [986.0, 984.0, 970.0] 80% band: [(844.0, 1148.0), (804.0, 1183.0), (759.0, 1191.0)]
condoms Aug-Oct 2021 median: [1365.0, 1351.0, 1333.0] 80% band: [(1080.0, 1609.0), (976.0, 1669.0), (855.0, 1713.0)]

Input. One 1-D float tensor per series: the monthly counts before the forecast origin on a regular monthly grid, oldest first, NaN for months that were not reported. The model was evaluated with 36 months of context (shorter series are fine). Pass all series of one sub-district in one call with cross_learning=True; series in the same call share information.

Output. predict_quantiles returns one tensor per input series of shape (1, prediction_length, n_quantiles) (the leading 1 is the target dimension) and the median. Counts can come out slightly negative near zero; clip at 0 if needed.

Helper for long tables

hmis_chronos.py (in this repository) does the windowing and grouping for a long table with one row per (series, month):

import sys
from huggingface_hub import snapshot_download
sys.path.insert(0, snapshot_download("judicialmind/HMIS-Chronos"))
from hmis_chronos import load, forecast

pipe = load("judicialmind/HMIS-Chronos")                   # bfloat16 on GPU, float32 on CPU
three = df[df.subdistrict.isin([1112, 1113, 1114])]         # df: series_id, ts, target, subdistrict (as above)
fc = forecast(pipe, three, origin="2021-08-01", horizon=3)
print(fc[["series_id", "h", "q0.1", "q0.5", "q0.9"]].head(3).round(1).to_string(index=False))

                 series_id  h  q0.1  q0.5   q0.9
hmis|10|188|1112|1.1|total  1 843.6 986.4 1148.2
hmis|10|188|1112|1.1|total  2 803.8 984.5 1182.8
hmis|10|188|1112|1.1|total  3 758.6 970.3 1191.1

forecast groups by subdistrict (set group_col=None to forecast each series alone), uses the 36 months before origin and skips series with fewer than 8 reported months in them, which is the HMIS-Bench window rule. The same module runs from the shell:

python hmis_chronos.py my_series.parquet --origin 2021-08-01 --horizon 3 --out forecasts.csv

Its forecasts match the HMIS-Bench forecast files of this model (computed on GPU in bfloat16) to a median relative difference of 0.07% on CPU in float32.

Files

file	content
`model.safetensors`, `config.json`	Chronos-2 weights and configuration (12 layers, d_model 768, 21 quantile heads)
`hmis_chronos.py`	`load`, `forecast` and a CLI for long monthly tables
`requirements.txt`	pinned runtime dependencies

Results

HMIS-Bench, averaged over six monthly origins (August 2021 to January 2022), on the same 599,714 windows for every model. chronos2_ft_hmis_s0 is this model. Lower WQL and MASE are better; a calibrated 80% interval covers 0.80; win rate = share of windows with lower MASE than zero-shot Chronos-2.

All series:

h	model	WQL	MASE	cov_80	win rate vs chronos2
1	HMIS-Chronos (`chronos2_ft_hmis_s0`)	0.1553	0.8224	0.8019	0.4985
1	`chronos2` (zero-shot)	0.1458	0.8380	0.8267	–
1	`tirex`	0.1435	0.8348	0.8278	0.5175
1	`timesfm25`	0.1459	0.8378	0.7568	0.5193
1	`lightgbm`	0.1473	0.8349	0.7989	0.5070
1	`naive`	0.1540	0.9081	0.7441	0.4833
2	HMIS-Chronos (`chronos2_ft_hmis_s0`)	0.1671	0.8738	0.8013	0.5073
2	`chronos2` (zero-shot)	0.1594	0.8949	0.8363	–
2	`tirex`	0.1575	0.8893	0.8404	0.5222
2	`timesfm25`	0.1582	0.8944	0.7521	0.5217
2	`lightgbm`	0.1574	0.8917	0.7959	0.5050
2	`naive`	0.1709	0.9823	0.7683	0.4674
3	HMIS-Chronos (`chronos2_ft_hmis_s0`)	0.1770	0.9153	0.8022	0.5094
3	`chronos2` (zero-shot)	0.1683	0.9376	0.8409	–
3	`tirex`	0.1665	0.9291	0.8476	0.5305
3	`timesfm25`	0.1662	0.9359	0.7472	0.5247
3	`lightgbm`	0.1634	0.9326	0.7923	0.5099
3	`naive`	0.1821	1.0405	0.7854	0.4570

Held-out states (Uttar Pradesh, Bihar, Odisha, Andhra Pradesh; not seen in fine-tuning):

h	model	WQL	MASE	cov_80	win rate vs chronos2
1	HMIS-Chronos	0.1740	0.8543	0.8112	0.4782
1	`chronos2` (zero-shot)	0.1698	0.8530	0.8335	–
2	HMIS-Chronos	0.1888	0.9248	0.8069	0.4816
2	`chronos2` (zero-shot)	0.1862	0.9252	0.8395	–
3	HMIS-Chronos	0.2048	0.9848	0.8045	0.4804
3	`chronos2` (zero-shot)	0.2006	0.9838	0.8410	–

HMIS-Chronos gives the most accurate median forecasts on the states it was tuned on (MASE 0.8224 vs 0.8380 at one month); its quantile bands are narrower than zero-shot Chronos-2, so WQL and 80% coverage favour the zero-shot model. On the held-out states the two are level on MASE. For calibrated intervals use zero-shot Chronos-2 or widen the HMIS-Chronos band; for point forecasts in the states covered by HMIS-Bench use HMIS-Chronos. The full leaderboard, with per-indicator slices, is in the HMIS-Bench-Leaderboard Space.

Coverage

Fine-tuned on HMIS-Bench sub-district series from April 2020 to July 2021 for 48 headline HMIS indicators in 28 States/UTs (Uttar Pradesh, Bihar, Odisha and Andhra Pradesh held out). Monthly frequency; forecasts 1 to 3 months.

Source data

Fine-tuned and evaluated on HMIS-Bench, derived from the HMIS sub-district monthly reports (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.

Base model: amazon/chronos-2 (Apache-2.0).

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{judicialmind2026hmischronos,
  title={HMIS-Chronos: Chronos-2 Fine-tuned for Indian Sub-district Health Service Forecasting},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/judicialmind/HMIS-Chronos}}
}

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.

Base model(s): amazon/chronos-2 (apache-2.0). Their licence terms and notices continue to apply to the upstream components; see THIRD_PARTY_NOTICES.md.