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.

BharatTS-Large

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 probabilistic forecaster for Indian open government time series (data.gov.in, NITI Aayog NDAP and other government portals): daily, weekly, monthly, quarterly and yearly series such as mandi prices, HMIS health indicators, rainfall, reservoir levels, crop yields and scheme statistics. It is one model with one entry point. Inside, a gate weighs a set of expert forecasters for each series, the way a mixture-of-experts layer does: eight Chronos-2 experts fine-tuned on BharatTS-Corpus, the base Chronos-2 and TimesFM-2.5 models, and statistical forecasters. You never pick components.

Headline result: on held-out series it never saw in training, every BharatTS-Large variant beats Chronos-2, TimesFM-2.5 and seasonal-naive on overall WQL.

Held-out series, overall WQL (lower is better; n = 8,349)

Model	WQL	GM-MASE vs seasonal-naive
BharatTS-Large v1 (frequency route)	0.171	0.665
BharatTS-Large distilled (single checkpoint)	0.175	0.662
BharatTS-Large v2 (full gate, default)	0.184	0.670
Chronos-2	0.207	0.716
TimesFM-2.5	0.257	0.715
Seasonal-naive	0.246	1.000

This release ships the v2 gate (router.json version 2) as the default and the distilled single-checkpoint variant.

Base models: amazon/chronos-2 (Apache-2.0) and google/timesfm-2.5-200m-pytorch (Apache-2.0, bundled unmodified). Their notices are kept in NOTICE.

Quickstart

hf download judicialmind/BharatTS-Large --local-dir BharatTS-Large   # gated: accept the terms, then `hf auth login`
pip install -r BharatTS-Large/requirements.txt

import sys
sys.path.insert(0, "BharatTS-Large")          # folder with the downloaded repo (bharatts_large.py lives there)

import numpy as np
from bharatts_large import BharatTSLarge

m = BharatTSLarge.from_pretrained("BharatTS-Large")   # local folder, or the Hub id "judicialmind/BharatTS-Large"

t = np.arange(72)
monthly = 100 + 10 * np.sin(2 * np.pi * t / 12) + 0.5 * t   # 6 years of monthly values, oldest first
monthly[30] = np.nan                                          # missing values are allowed
yearly = np.array([52.0, 55.1, 57.9, 61.2, 60.4, 64.8, 68.0, 71.3])

q = m.predict([monthly, monthly * 2], freq="M")               # default horizon for M is 6
print(q.shape)                                                # (2, 6, 9): series x horizon x quantiles 0.1..0.9
print(np.round(q[0, :, 4], 1))                                # median forecast of series 0

q = m.predict([yearly], freq="Y", horizon=5, quantiles=[0.05, 0.5, 0.95])
print(q.shape)                                                # (1, 5, 3)
print(np.round(q[0], 1))                                      # rows = years ahead, columns = 5% / 50% / 95%

Output (CPU): the first forecast has shape (2, 6, 9) (series × horizon × quantiles) and the second (1, 5, 3).

Median monthly forecast, series 0

Months ahead	1	2	3	4	5	6
Median	135.6	141.5	146.0	147.8	147.2	143.7

Yearly forecast with 5% / 50% / 95% quantiles

Years ahead	5%	50%	95%
1	68.5	73.7	79.1
2	68.1	75.8	84.3
3	67.9	78.1	89.2
4	67.7	77.8	88.6
5	68.0	78.7	90.7

Input. contexts: a list of 1-D numeric arrays (lists, numpy arrays or tensors), one per series, oldest value first, NaN for missing values, any length (the last 2048 points are used). freq: one frequency code for the whole batch, "D", "W", "M", "Q" or "Y". Optional horizon (defaults: D 30, W 13, M 6, Q 4, Y 3) and quantiles (default 0.1, 0.2, ..., 0.9; other levels are interpolated and clamped to the 0.1 / 0.9 forecasts outside that range).

Output. A numpy.float64 array of shape (n_series, horizon, n_quantiles), in the units of the input; quantiles never cross. Take [..., 4] of the default output for the median.

from_pretrained also accepts device="cuda" / "cpu" (default: CUDA if available; experts run in bf16 on GPU, fp32 on CPU) and batch_size. Experts are loaded on first use; loaded by Hub id, their files are downloaded on first use.

Variant: `distilled`

m = BharatTSLarge.from_pretrained("BharatTS-Large", variant="distilled")   # default (variant=None) is the full gate
q = m.predict([monthly], freq="M")

A single Chronos-2 checkpoint distilled from the gate: one forward pass per series, ~120M parameters.

Files

file	content
`bharatts_large.py`	the runtime: `BharatTSLarge.from_pretrained(...)`, `.predict(...)`, statistical forecasters, gate
`router.json`	frozen gate configuration: horizons and seasons per frequency, frequency routes, local-backtest rules, expert list
`experts/*.safetensors`	Chronos-2 experts (bf16); `experts/chronos2_config_*.json` are their shared configs
`experts/timesfm25/`	google/timesfm-2.5-200m-pytorch, unmodified
`eval/leaderboard.md`, `eval/leaderboard.json`	test results, overall and by ministry
`variants/distilled.safetensors`	the distilled single-checkpoint variant (Chronos-2 architecture, bf16; config in `experts/`)
`eval/distilled/`	test results of the distilled variant (same windows): `leaderboard.md`, `leaderboard.json`, `by_freq.parquet`, `by_ministry.parquet`
`eval/by_freq.parquet`, `eval/by_ministry.parquet`	test results per frequency / per ministry (`task`, `freq` or `ministry`, `model`, `n`, `mase`, `wql`, `gm_rel_mase_vs_snaive`)
`requirements.txt`	pinned runtime dependencies
`NOTICE`	Apache-2.0 notices of the base models

Experts

expert	what it is	weights
`bharatts_g1` ... `bharatts_g5`	BharatTS fine-tunes of amazon/chronos-2 on all frequencies	`experts/bharatts_g1.safetensors` ... `experts/bharatts_g5.safetensors`
`bharatts_w`	BharatTS fine-tune of amazon/chronos-2, weekly specialist	`experts/bharatts_w.safetensors`
`bharatts_q`	BharatTS fine-tune of amazon/chronos-2, quarterly specialist	`experts/bharatts_q.safetensors`
`bharatts_y`	BharatTS fine-tune of amazon/chronos-2, yearly specialist	`experts/bharatts_y.safetensors`
`chronos2`	amazon/chronos-2	`experts/chronos2.safetensors`
`timesfm25`	google/timesfm-2.5-200m-pytorch (unmodified)	`experts/timesfm25`
`seasonal_naive`, `naive`, `drift`, `mean`, `ets_lite`	statistical forecasters with empirical residual quantiles	computed in `bharatts_large.py`

Chronos-2 experts are stored as bf16 safetensors; experts with identical configs share one config file. GPU inference runs in bf16, so this loses nothing there; on CPU the weights are upcast to fp32.

Gate

Frequency route (router): each frequency has a fixed quantile blend of experts.

freq	router
D	1/3 `bharatts_g4` + 1/3 `bharatts_g1` + 1/3 `bharatts_g3`
W	0.75 `bharatts_w` + 0.25 `seasonal_naive`
M	0.5 `bharatts_g5` + 0.5 `bharatts_g2`
Q	`bharatts_q`
Y	0.75 `bharatts_g1` + 0.25 `bharatts_y`

Per-series local backtest: for each series the candidates forecast that series' own most recent k windows inside the context you pass (never beyond it). Each candidate gets a scale-free score (pinball loss / sum |y|). The forecast is the best candidate if it beats router by the margin, otherwise router. A series too short for any backtest window gets router.

freq	k	rule	candidates
D	2	best, margin 0.5	router, chronos2, timesfm25, seasonal_naive, router+chronos2, router+timesfm25, router+chronos2+timesfm25, router+seasonal_naive
W	5	best, margin 0.1	router, chronos2, timesfm25, seasonal_naive, router+chronos2, router+timesfm25, router+chronos2+timesfm25, router+seasonal_naive
M	5	best, margin 0.1	router, chronos2, timesfm25, seasonal_naive, router+chronos2, router+timesfm25, router+chronos2+timesfm25, router+seasonal_naive
Q	3	best, margin 0.25	router, chronos2, timesfm25, seasonal_naive, router+chronos2, router+timesfm25, router+chronos2+timesfm25, router+seasonal_naive
Y	1	best, margin 1.0	router, chronos2, timesfm25, seasonal_naive

a+b candidates are equal-weight quantile blends. The gate was frozen before the test windows were scored.

Results

Test windows, one pass of the frozen gate (eval/leaderboard.json, eval/by_freq.parquet). Lower is better for both metrics. WQL = pooled weighted quantile loss; GM-MASE = geometric mean of the MASE ratio to seasonal-naive. v2 is the full gate (default); v1 is its frequency route alone (step 1 of the gate); distilled is the single-checkpoint variant. heldout = series never trained on, last h points; time = training series, the h points after the per-frequency calendar cutoff.

Overall test results (eval/leaderboard.json, eval/distilled/leaderboard.json)

Split	n	Metric	v2 (default)	distilled	v1	Chronos-2	TimesFM-2.5	Seasonal-naive
heldout	8,349	WQL	0.184	0.175	0.171	0.207	0.257	0.246
heldout	8,349	GM-MASE	0.670	0.662	0.665	0.716	0.715	1.000
time	5,932	WQL	0.188	0.191	0.187	0.195	0.202	0.276
time	5,932	GM-MASE	0.614	0.630	0.619	0.647	0.635	1.000

Held-out series by frequency, WQL (GM-MASE) — frequencies where a BharatTS-Large variant is best

Freq	n	v2 (default)	distilled	v1	Chronos-2	TimesFM-2.5	Seasonal-naive
D	2,303	0.049 (0.606)	0.050 (0.593)	0.053 (0.598)	0.054 (0.658)	0.059 (0.635)	0.091 (1.000)
M	3,208	0.184 (0.556)	0.175 (0.564)	0.171 (0.549)	0.207 (0.601)	0.257 (0.588)	0.246 (1.000)
Q	300	0.444 (0.717)	0.333 (0.591)	0.444 (0.711)	0.396 (0.745)	0.428 (0.696)	0.973 (1.000)

All frequencies and the per-ministry breakdown: eval/by_freq.parquet, eval/leaderboard.md, eval/by_ministry.parquet (distilled: eval/distilled/).

Coverage

Indian government statistics at national, state, district and sub-district level, D/W/M/Q/Y frequencies, series starting between 1901 and 2026, from 8,461 source tables (1,799,285 series in BharatTS-Corpus). The heldout series of BharatTS-Corpus were never trained on.

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.

Expert weights derived from or equal to Apache-2.0 models (amazon/chronos-2, google/timesfm-2.5-200m-pytorch); see NOTICE.

Original data: published by the Indian ministries, departments and agencies named per series in the ministry column of BharatTS-Corpus 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{judicialmind2026bharattslarge,
  title        = {BharatTS-Large: A Mixture-of-Experts Forecaster for Indian Open Government Time Series},
  author       = {JudicialMind},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/judicialmind/BharatTS-Large}}
}

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), google/timesfm-2.5-200m-pytorch (apache-2.0). Their licence terms and notices continue to apply to the upstream components; see THIRD_PARTY_NOTICES.md and NOTICE.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Time Series Forecasting

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for judicialmind/BharatTS-Large

Base model

amazon/chronos-2

Finetuned

(13)

this model

Collection including judicialmind/BharatTS-Large

India Open Government Data — Bharat Forecasting Models

Collection

Research and non-commercial use (CC BY-NC-SA 4.0). Commercial: surya@judicialmind.ai / hello@judicialmind.ai • 2 items • Updated 1 day ago