Datasets:

judicialmind
/

CropBench

---
license: cc-by-nc-sa-4.0
task_categories:
- time-series-forecasting
tags:
- agriculture
- india
- crop-yield
- forecasting
- benchmark
pretty_name: CropBench
language:
- en
size_categories:
- 100K<n<1M
configs:
- config_name: series
  data_files: series/series.parquet
- config_name: panel
  data_files: panel/panel.parquet
- config_name: covariates
  data_files: series/covariates.parquet
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
---
# CropBench

<!-- 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 -->

CropBench is a forecasting benchmark for district crop yield and production in India, built from official district
crop statistics. Each series is one district x crop x season (Kharif, Rabi, Whole Year, Summer, Winter, Autumn) over
the agricultural years 1997-2019, with rainfall, mandi price and reservoir covariates. Models forecast the next season
from five rolling origins (2015-2019), with or without covariates. The repository ships the panel, the benchmark
series and covariates, the forecasts of eight models, the leaderboard and `evaluate.py`, which reproduces
`leaderboard.csv` from `forecasts/` and scores new models.

## Quickstart

```python
from huggingface_hub import snapshot_download
root = snapshot_download("judicialmind/CropBench", repo_type="dataset")
import pandas as pd
series = pd.read_parquet(f"{root}/series/series.parquet")
print(series.shape, series["series_id"].nunique(), "series")
print(series[["series_id", "year", "area_ha", "production", "yield"]].head(3).to_string(index=False))
```

Output: `(209280, 15) 12840 series` (209,280 rows, 15 columns, 12,840 series), then:

**First three rows of `series/series.parquet`**

| series id | year | area (ha) | production | yield |
|---|---:|---:|---:|---:|
| 1\|Dry chillies\|kharif | 2005 | 53.0 | 62.8 | 1.1849 |
| 1\|Dry chillies\|kharif | 2006 | 41.0 | 48.6 | 1.1854 |
| 1\|Dry chillies\|kharif | 2007 | 78.0 | 92.5 | 1.1859 |

Other loaders: `datasets.load_dataset("judicialmind/CropBench", "series", split="train")` (also `panel`,
`covariates`), or DuckDB (`SELECT crop, avg(yield) FROM 'panel/panel.parquet' GROUP BY 1`).

### Evaluate a model

Input: for each benchmark row (`series_id, origin, target`; `target` is `yield` or `production`) the model may use the
series' values for years before `origin` (and, for covariate models, the covariate rows of the chosen setting).
Output: nine quantiles `q0.1 ... q0.9` of the target in year `origin`. `evaluate.py` (numpy, pandas, pyarrow) scores
them:

```python
import sys
import numpy as np
sys.path.insert(0, root)
from evaluate import QCOLS, QUANTILES, CropBench, leaderboard
bench = CropBench(root)
truth = bench.truth()            # the rows to forecast: series_id, origin, target (+ ground truth, never use it)
# toy forecaster: mean of the last 3 observed years before the origin, +-25% normal spread
hist = series.melt(id_vars=["series_id", "year"], value_vars=["yield", "production"], var_name="target", value_name="v")
rows = truth[["series_id", "origin", "target"]].merge(hist, on=["series_id", "target"])
rows = rows[(rows["year"] < rows["origin"]) & rows["v"].notna()].sort_values("year")
point = rows.groupby(["series_id", "origin", "target"])["v"].apply(lambda v: v.tail(3).mean()).rename("point")
sub = point.reset_index()
z = np.array([-1.2816, -0.8416, -0.5244, -0.2533, 0.0, 0.2533, 0.5244, 0.8416, 1.2816])
sub[QCOLS] = sub["point"].to_numpy()[:, None] * (1 + 0.25 * z)
sub["setting"] = "univariate"
sub.to_parquet("last3_mean.parquet")    # submission file: series_id, origin, target, setting, q0.1 ... q0.9
fc = truth.merge(sub, on=["series_id", "origin", "target"]).assign(model="last3_mean")
lb = leaderboard(fc)
print(lb[lb["setting"] == "ex_ante"][["target", "model", "swql", "wql", "mase", "mdase", "cov_80", "n"]]
      .round(4).to_string(index=False))
```

**Toy forecaster `last3_mean`, ex-ante setting**
| target | model | sWQL | WQL | MASE | MdASE | 80% coverage | rows |
|---|---|---:|---:|---:|---:|---:|---:|
| production | last3_mean | 0.4569 | 0.1577 | 1.7572 | 0.7218 | 0.5161 | 32,599 |
| yield | last3_mean | 0.2078 | 0.1612 | 2.3760 | 0.8292 | 0.7323 | 32,599 |
From the command line (run inside the downloaded folder):
```bash
python evaluate.py                     # recompute leaderboard.csv (and the per-origin / per-crop rows) -> "REPRODUCED"
python evaluate.py --submission last3_mean.parquet --name last3_mean --out last3_mean.csv   # ranked with the shipped models
```
A submission has one row per benchmark row and setting: `series_id, origin, target, setting, q0.1 ... q0.9`, with
`setting` = `ex_ante`, `in_season`, or `univariate` (no covariates; listed under both settings). To add a model to
the leaderboard, open a discussion on this repository with the `--out` csv and a short model description.
## Tasks
- **Targets:** `yield` (production / area) and `production`.
- **Origins:** for origin year Y the model sees each series' values for years before Y (context up to
  23 years, at least 8 observed) and forecasts year Y.
- **Settings:** `ex_ante` (information dated before sowing: pre-season rainfall, last year's season rainfall,
  pre-sowing mandi price change, reservoir storage) and `in_season` (adds the season's observed rainfall; official
  district estimates are published one to two years later, so this is still a forecast of an unpublished number).
  Univariate models are setting-free and shown under both.
- **No leakage:** context timestamps precede the origin; every covariate carries the last date it depends on
  (`cov_max_date`), which precedes the row's cut-off; LightGBM is retrained per origin on target years before the
  origin. Rainfall anomalies use a 1951-1996 climatology (before the first panel year).
- **Metrics:** sWQL is the headline metric: the weighted quantile loss (9 quantiles) after each test row's target
  and quantiles are divided by that series' mean over its context years, so every series is on the same scale.
  The plain pooled WQL is also shown but should not be used to rank models here: units differ by crop (coconut
  production is counted in nuts, cotton in bales), and Coconut is 96% of the summed absolute yield values and of the
  summed absolute production values in the test rows, so pooled WQL mostly measures that crop. Also shown: MASE
  (scale = mean absolute year-on-year change in the context) and its median MdASE, MAPE and median APE of the median
  forecast, and 80% interval coverage. Means of scaled and percentage errors are dominated by a few series
  (near-flat history, collapsing production), so the medians are shown next to them. Rows are sorted by sWQL.
## Results
Leaderboard pooled over origins (`leaderboard.csv`; per origin and per crop in `leaderboard.json`).
### target = production, setting = ex_ante
| model | covariates | sWQL | WQL (pooled) | MASE | MdASE | MAPE | MdAPE | cov_80 | n |
|---|---|---|---|---|---|---|---|---|---|
| lightgbm | no | 0.4086 | 0.1367 | 1.656 | 0.595 | 2.214 | 0.276 | 0.749 | 32,599 |
| lightgbm_cov | yes | 0.4096 | 0.1384 | 1.655 | 0.594 | 2.222 | 0.272 | 0.749 | 32,599 |
| chronos2 | no | 0.4241 | 0.1411 | 1.707 | 0.630 | 2.964 | 0.290 | 0.777 | 32,599 |
| chronos2_cov | yes | 0.4344 | 0.1419 | 1.769 | 0.709 | 3.665 | 0.311 | 0.773 | 32,599 |
| naive | no | 0.4515 | 0.1503 | 1.676 | 0.583 | 2.175 | 0.270 | 0.693 | 32,599 |
| ets_lite | no | 0.4566 | 0.1514 | 1.719 | 0.686 | 3.354 | 0.299 | 0.625 | 32,599 |
| mean3 | no | 0.4663 | 0.1591 | 1.757 | 0.722 | 3.377 | 0.300 | 0.675 | 32,599 |
| hist_mean | no | 0.5863 | 0.2106 | 2.520 | 1.353 | 10.545 | 0.498 | 0.453 | 32,599 |
### target = production, setting = in_season
| model | covariates | sWQL | WQL (pooled) | MASE | MdASE | MAPE | MdAPE | cov_80 | n |
|---|---|---|---|---|---|---|---|---|---|
| lightgbm_cov | yes | 0.4082 | 0.1361 | 1.650 | 0.593 | 2.176 | 0.270 | 0.751 | 32,599 |
| lightgbm | no | 0.4086 | 0.1367 | 1.656 | 0.595 | 2.214 | 0.276 | 0.749 | 32,599 |
| chronos2 | no | 0.4241 | 0.1411 | 1.707 | 0.630 | 2.964 | 0.290 | 0.777 | 32,599 |
| chronos2_cov | yes | 0.4330 | 0.1411 | 1.760 | 0.687 | 3.540 | 0.305 | 0.755 | 32,599 |
| naive | no | 0.4515 | 0.1503 | 1.676 | 0.583 | 2.175 | 0.270 | 0.693 | 32,599 |
| ets_lite | no | 0.4566 | 0.1514 | 1.719 | 0.686 | 3.354 | 0.299 | 0.625 | 32,599 |
| mean3 | no | 0.4663 | 0.1591 | 1.757 | 0.722 | 3.377 | 0.300 | 0.675 | 32,599 |
| hist_mean | no | 0.5863 | 0.2106 | 2.520 | 1.353 | 10.545 | 0.498 | 0.453 | 32,599 |
### target = yield, setting = ex_ante
| model | covariates | sWQL | WQL (pooled) | MASE | MdASE | MAPE | MdAPE | cov_80 | n |
|---|---|---|---|---|---|---|---|---|---|
| lightgbm_cov | yes | 0.1849 | 0.1532 | 2.346 | 0.767 | 0.296 | 0.128 | 0.743 | 32,599 |
| lightgbm | no | 0.1849 | 0.1530 | 2.348 | 0.764 | 0.298 | 0.129 | 0.750 | 32,599 |
| chronos2 | no | 0.1953 | 0.1517 | 2.349 | 0.767 | 0.285 | 0.138 | 0.789 | 32,599 |
| chronos2_cov | yes | 0.1978 | 0.1520 | 2.385 | 0.821 | 0.288 | 0.143 | 0.762 | 32,599 |
| ets_lite | no | 0.2072 | 0.1611 | 2.344 | 0.780 | 0.309 | 0.134 | 0.600 | 32,599 |
| mean3 | no | 0.2098 | 0.1661 | 2.376 | 0.829 | 0.307 | 0.142 | 0.637 | 32,599 |
| naive | no | 0.2149 | 0.1690 | 2.350 | 0.740 | 0.310 | 0.130 | 0.661 | 32,599 |
| hist_mean | no | 0.2494 | 0.2174 | 2.966 | 1.203 | 0.339 | 0.209 | 0.533 | 32,599 |
### target = yield, setting = in_season
| model | covariates | sWQL | WQL (pooled) | MASE | MdASE | MAPE | MdAPE | cov_80 | n |
|---|---|---|---|---|---|---|---|---|---|
| lightgbm_cov | yes | 0.1829 | 0.1525 | 2.330 | 0.760 | 0.294 | 0.127 | 0.745 | 32,599 |
| lightgbm | no | 0.1849 | 0.1530 | 2.348 | 0.764 | 0.298 | 0.129 | 0.750 | 32,599 |
| chronos2 | no | 0.1953 | 0.1517 | 2.349 | 0.767 | 0.285 | 0.138 | 0.789 | 32,599 |
| chronos2_cov | yes | 0.1973 | 0.1531 | 2.385 | 0.812 | 0.286 | 0.142 | 0.738 | 32,599 |
| ets_lite | no | 0.2072 | 0.1611 | 2.344 | 0.780 | 0.309 | 0.134 | 0.600 | 32,599 |
| mean3 | no | 0.2098 | 0.1661 | 2.376 | 0.829 | 0.307 | 0.142 | 0.637 | 32,599 |
| naive | no | 0.2149 | 0.1690 | 2.350 | 0.740 | 0.310 | 0.130 | 0.661 | 32,599 |
| hist_mean | no | 0.2494 | 0.2174 | 2.966 | 1.203 | 0.339 | 0.209 | 0.533 | 32,599 |
Models: `naive` (last observed value), `mean3` (mean of the last three observed years), `hist_mean` (mean of the
context), `ets_lite` (damped-trend exponential smoothing), `lightgbm` / `lightgbm_cov` (global LightGBM quantile
regression on scale-free history features, without / with the setting's covariates, retrained per origin),
`chronos2` / `chronos2_cov` (`amazon/chronos-2` zero-shot, without / with the covariates).
## Files
| file | content |
|---|---|
| `panel/panel.parquet` | district x crop x season x year area, production, yield (297,505 rows) |
| `series/series.parquet` | the 12,840 benchmark series, one row per series and year (209,280 rows) |
| `series/covariates.parquet` | covariates per series-year and setting (418,560 rows) |
| `splits.json` | origins, context length (23), minimum context (8), windows per origin |
| `forecasts/<model>/<target>_<setting>_<origin>.parquet` | forecasts of the eight models |
| `leaderboard.csv`, `leaderboard.md` | the leaderboard above (32 rows) |
| `leaderboard.json` | the leaderboard plus per-origin (100) and per-crop (1,080) rows and the crop shares of the pooled targets |
| `sources/imd_district_resources.csv` | the 867 IMD gridded district rainfall resources (one per month, 1951-2023) |
| `evaluate.py`, `requirements.txt` | evaluation harness |
## Schema
`panel/panel.parquet` (`series/series.parquet` has the same columns plus `series_id`, `ts`, `target`)
| column | type | description |
|---|---|---|
| `state_lgd`, `state_name` | int16, string | state (LGD code and name) |
| `district_lgd`, `district_name` | int32, string | district (LGD code and name) |
| `crop` | string | crop name (56 crops) |
| `season` | string | `kharif`, `rabi`, `whole_year`, `summer`, `winter`, `autumn` |
| `year` | int16 | agricultural year (July-June) starting in this year |
| `area_ha` | float64 | area, hectares |
| `production` | float64 | production: tonnes; bales for cotton, jute and mesta; nuts for coconut |
| `yield` | float64 | production / area |
| `src` | string | `ndap` (NDAP dataset 6820) or `ogd_apy` (data.gov.in resource 35be999b..., filling keys missing from NDAP) |
| `yield_reported` | float64 | yield as published in the source |
| `series_id` | string | `district_lgd|crop|season` (series file only) |
| `ts` | date | July 1 of `year` (series file only) |
| `target` | float64 | equal to `yield` (series file only) |
`series/covariates.parquet`: keys `district_lgd, crop, season, year, setting` (`ex_ante` / `in_season`) and
| column | description |
|---|---|
| `cov_rain_season` | log((season rain + 10 mm) / (1951-1996 climatology + 10 mm)); `in_season` only |
| `cov_rain_pre` | the same for the pre-season months before sowing |
| `cov_rain_prev` | last year's season rainfall anomaly |
| `cov_price_pre` | log(median mandi modal price of the crop in the 3 months before sowing / the same months a year earlier); district level, state median as fallback |
| `cov_reservoir` | national live storage / live capacity of CWC reservoirs on the last report before the cut-off (at most 21 days old) |
| `cov_rain_tier` | 1 = district IMD gridded rainfall, 2 = state mean of IMD sub-divisions, 0 = none |
| `cov_max_date` | latest date any covariate value depends on |
| `cutoff` | the row's information cut-off: the season's sowing month for `ex_ante`, the month after the season's rainfall months for `in_season` |
`forecasts/<model>/*.parquet`: `series_id, district_lgd, state_lgd, crop, season, origin, h` (= 1), `ts, q0.1 ...
q0.9, mean`, `target_value` (the observed value), `mase_scale`, `model`, `target`, `setting` (`univariate`,
`ex_ante` or `in_season`).
`sources/imd_district_resources.csv`: `uuid` (data.gov.in resource id), `month`, `api_total` (rows the API reports),
`records_in_dump` (rows in the snapshot).
## Coverage
| | |
|---|---|
| panel rows (district x crop x season x year, with yield) | 292,333 |
| districts / states / crops | 671 / 35 / 56 |
| benchmark series (median area >= 50 ha, >= 8 years) | 12,840 |
| rolling origins (agricultural years) | 2015, 2016, 2017, 2018, 2019 |
| horizon | 1 year (next season of the same type) |
| test rows per target | 32,599 (6,984 / 7,248 / 7,343 / 7,357 / 3,667 for 2015-2019) |
Rainfall: district-level (tier 1) rainfall covers 340 districts; the others use state-level sub-division means
(tier 2). Share of benchmark series-years by tier: {'1': 0.4168, '2': 0.5627, '0': 0.0205}.
- IMD gridded district rainfall: every monthly file in the snapshot holds only the first 10,000 of about 22,700 API
  rows, so only districts of states sorted first (Andaman & Nicobar .. part of Madhya Pradesh) have district
  rainfall; other districts use the state mean of IMD sub-divisions (to 2017).
- NDAP indicator columns are unlabeled in the dump; area/production/yield were identified by the identity
  yield = production / area (holds on 100.0% of rows).
- Agricultural year 2019-20 is partial in the source (about half the districts).
- PM-KISAN is not a covariate: instalments start in 2019, after the last complete panel year.
- About 3% of panel values (4-6% of Kharif, Rabi and Whole Year test rows) repeat the previous year's figure exactly.
## Source data
Derived from 871 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).
Providers: Central Water Commission, India Meteorological Department, Ministry of Agriculture and Farmers Welfare, NITI Aayog (NDAP).
- Source: NITI Aayog, National Data and Analytics Platform (NDAP), dataset 6820: Land Use Statistics : Crop Production Statistics (Ministry of Agriculture and Farmers Welfare). https://ndap.niti.gov.in/dataset/6820
- Department of Agriculture and Farmers Welfare (Ministry of Agriculture and Farmers Welfare). "District-wise, season-wise crop production statistics from 1997". data.gov.in resource 35be999b-0208-4354-b557-f6ca9a5355de, https://api.data.gov.in/resource/35be999b-0208-4354-b557-f6ca9a5355de. Licensed under the Government Open Data License - India (https://data.gov.in/government-open-data-license-india). Snapshot: May 2026.
- India Meteorological Department (IMD) (Ministry of Earth Sciences). "Sub Divisional Monthly Rainfall from 1901 to 2017". data.gov.in resource 8e0bd482-4aba-4d99-9cb9-ff124f6f1c2f, https://api.data.gov.in/resource/8e0bd482-4aba-4d99-9cb9-ff124f6f1c2f. Licensed under the Government Open Data License - India (https://data.gov.in/government-open-data-license-india). Snapshot: May 2026.
- IMD GRID MODEL daily district rainfall, one data.gov.in resource per month (resource list in `sources/imd_district_resources.csv`). Licensed under the Government Open Data License - India.
- Agmarknet daily mandi prices (data.gov.in resource 35985678-0d79-46b4-9ed6-6f13308a1d24) and CWC reservoir storage (data.gov.in resource 1fc2148c-fc41-46f5-a364-bdc03f77053f). Licensed under the Government Open Data License - India.
GODL-India does not cover personal information; this dataset holds aggregate statistics only.
<!-- jm:source -->
Original data: published by Department of Agriculture and Farmers Welfare, Ministry of Agriculture and Farmers Welfare (district crop production statistics, via data.gov.in and NITI Aayog NDAP); India Meteorological Department, Ministry of Earth Sciences (rainfall); Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare (mandi prices); Central Water Commission, Ministry of Jal Shakti (reservoir storage) 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{judicialmind2026cropbench,
  title        = {CropBench: a forecasting benchmark for district crop yield and production in India},
  author       = {JudicialMind},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/judicialmind/CropBench}}
}
```
<!-- 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 -->

	---
	license: cc-by-nc-sa-4.0
	task_categories:
	- time-series-forecasting
	tags:
	- agriculture
	- india
	- crop-yield
	- forecasting
	- benchmark
	pretty_name: CropBench
	language:
	- en
	size_categories:
	- 100K<n<1M
	configs:
	- config_name: series
	data_files: series/series.parquet
	- config_name: panel
	data_files: panel/panel.parquet
	- config_name: covariates
	data_files: series/covariates.parquet
	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
	---

	# CropBench

	<!-- 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 -->

	CropBench is a forecasting benchmark for district crop yield and production in India, built from official district
	crop statistics. Each series is one district x crop x season (Kharif, Rabi, Whole Year, Summer, Winter, Autumn) over
	the agricultural years 1997-2019, with rainfall, mandi price and reservoir covariates. Models forecast the next season
	from five rolling origins (2015-2019), with or without covariates. The repository ships the panel, the benchmark
	series and covariates, the forecasts of eight models, the leaderboard and `evaluate.py`, which reproduces
	`leaderboard.csv` from `forecasts/` and scores new models.

	## Quickstart

	```python
	from huggingface_hub import snapshot_download

	root = snapshot_download("judicialmind/CropBench", repo_type="dataset")

	import pandas as pd

	series = pd.read_parquet(f"{root}/series/series.parquet")
	print(series.shape, series["series_id"].nunique(), "series")
	print(series[["series_id", "year", "area_ha", "production", "yield"]].head(3).to_string(index=False))
	```

	Output: `(209280, 15) 12840 series` (209,280 rows, 15 columns, 12,840 series), then:

	First three rows of `series/series.parquet`

	\| series id \| year \| area (ha) \| production \| yield \|
	\|---\|---:\|---:\|---:\|---:\|
	\| 1\\|Dry chillies\\|kharif \| 2005 \| 53.0 \| 62.8 \| 1.1849 \|
	\| 1\\|Dry chillies\\|kharif \| 2006 \| 41.0 \| 48.6 \| 1.1854 \|
	\| 1\\|Dry chillies\\|kharif \| 2007 \| 78.0 \| 92.5 \| 1.1859 \|

	Other loaders: `datasets.load_dataset("judicialmind/CropBench", "series", split="train")` (also `panel`,
	`covariates`), or DuckDB (`SELECT crop, avg(yield) FROM 'panel/panel.parquet' GROUP BY 1`).

	### Evaluate a model

	Input: for each benchmark row (`series_id, origin, target`; `target` is `yield` or `production`) the model may use the
	series' values for years before `origin` (and, for covariate models, the covariate rows of the chosen setting).
	Output: nine quantiles `q0.1 ... q0.9` of the target in year `origin`. `evaluate.py` (numpy, pandas, pyarrow) scores
	them:

	```python
	import sys

	import numpy as np

	sys.path.insert(0, root)
	from evaluate import QCOLS, QUANTILES, CropBench, leaderboard

	bench = CropBench(root)
	truth = bench.truth() # the rows to forecast: series_id, origin, target (+ ground truth, never use it)

	# toy forecaster: mean of the last 3 observed years before the origin, +-25% normal spread
	hist = series.melt(id_vars=["series_id", "year"], value_vars=["yield", "production"], var_name="target", value_name="v")
	rows = truth[["series_id", "origin", "target"]].merge(hist, on=["series_id", "target"])
	rows = rows[(rows["year"] < rows["origin"]) & rows["v"].notna()].sort_values("year")
	point = rows.groupby(["series_id", "origin", "target"])["v"].apply(lambda v: v.tail(3).mean()).rename("point")
	sub = point.reset_index()
	z = np.array([-1.2816, -0.8416, -0.5244, -0.2533, 0.0, 0.2533, 0.5244, 0.8416, 1.2816])
	sub[QCOLS] = sub["point"].to_numpy()[:, None] * (1 + 0.25 * z)
	sub["setting"] = "univariate"
	sub.to_parquet("last3_mean.parquet") # submission file: series_id, origin, target, setting, q0.1 ... q0.9

	fc = truth.merge(sub, on=["series_id", "origin", "target"]).assign(model="last3_mean")
	lb = leaderboard(fc)
	print(lb[lb["setting"] == "ex_ante"][["target", "model", "swql", "wql", "mase", "mdase", "cov_80", "n"]]
	.round(4).to_string(index=False))
	```

	Toy forecaster `last3_mean`, ex-ante setting

	\| target \| model \| sWQL \| WQL \| MASE \| MdASE \| 80% coverage \| rows \|
	\|---\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| production \| last3_mean \| 0.4569 \| 0.1577 \| 1.7572 \| 0.7218 \| 0.5161 \| 32,599 \|
	\| yield \| last3_mean \| 0.2078 \| 0.1612 \| 2.3760 \| 0.8292 \| 0.7323 \| 32,599 \|

	From the command line (run inside the downloaded folder):

	```bash
	python evaluate.py # recompute leaderboard.csv (and the per-origin / per-crop rows) -> "REPRODUCED"
	python evaluate.py --submission last3_mean.parquet --name last3_mean --out last3_mean.csv # ranked with the shipped models
	```

	A submission has one row per benchmark row and setting: `series_id, origin, target, setting, q0.1 ... q0.9`, with
	`setting` = `ex_ante`, `in_season`, or `univariate` (no covariates; listed under both settings). To add a model to
	the leaderboard, open a discussion on this repository with the `--out` csv and a short model description.

	## Tasks

	- Targets: `yield` (production / area) and `production`.
	- Origins: for origin year Y the model sees each series' values for years before Y (context up to
	23 years, at least 8 observed) and forecasts year Y.
	- Settings: `ex_ante` (information dated before sowing: pre-season rainfall, last year's season rainfall,
	pre-sowing mandi price change, reservoir storage) and `in_season` (adds the season's observed rainfall; official
	district estimates are published one to two years later, so this is still a forecast of an unpublished number).
	Univariate models are setting-free and shown under both.
	- No leakage: context timestamps precede the origin; every covariate carries the last date it depends on
	(`cov_max_date`), which precedes the row's cut-off; LightGBM is retrained per origin on target years before the
	origin. Rainfall anomalies use a 1951-1996 climatology (before the first panel year).
	- Metrics: sWQL is the headline metric: the weighted quantile loss (9 quantiles) after each test row's target
	and quantiles are divided by that series' mean over its context years, so every series is on the same scale.
	The plain pooled WQL is also shown but should not be used to rank models here: units differ by crop (coconut
	production is counted in nuts, cotton in bales), and Coconut is 96% of the summed absolute yield values and of the
	summed absolute production values in the test rows, so pooled WQL mostly measures that crop. Also shown: MASE
	(scale = mean absolute year-on-year change in the context) and its median MdASE, MAPE and median APE of the median
	forecast, and 80% interval coverage. Means of scaled and percentage errors are dominated by a few series
	(near-flat history, collapsing production), so the medians are shown next to them. Rows are sorted by sWQL.

	## Results

	Leaderboard pooled over origins (`leaderboard.csv`; per origin and per crop in `leaderboard.json`).

	### target = production, setting = ex_ante

	\| model \| covariates \| sWQL \| WQL (pooled) \| MASE \| MdASE \| MAPE \| MdAPE \| cov_80 \| n \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| lightgbm \| no \| 0.4086 \| 0.1367 \| 1.656 \| 0.595 \| 2.214 \| 0.276 \| 0.749 \| 32,599 \|
	\| lightgbm_cov \| yes \| 0.4096 \| 0.1384 \| 1.655 \| 0.594 \| 2.222 \| 0.272 \| 0.749 \| 32,599 \|
	\| chronos2 \| no \| 0.4241 \| 0.1411 \| 1.707 \| 0.630 \| 2.964 \| 0.290 \| 0.777 \| 32,599 \|
	\| chronos2_cov \| yes \| 0.4344 \| 0.1419 \| 1.769 \| 0.709 \| 3.665 \| 0.311 \| 0.773 \| 32,599 \|
	\| naive \| no \| 0.4515 \| 0.1503 \| 1.676 \| 0.583 \| 2.175 \| 0.270 \| 0.693 \| 32,599 \|
	\| ets_lite \| no \| 0.4566 \| 0.1514 \| 1.719 \| 0.686 \| 3.354 \| 0.299 \| 0.625 \| 32,599 \|
	\| mean3 \| no \| 0.4663 \| 0.1591 \| 1.757 \| 0.722 \| 3.377 \| 0.300 \| 0.675 \| 32,599 \|
	\| hist_mean \| no \| 0.5863 \| 0.2106 \| 2.520 \| 1.353 \| 10.545 \| 0.498 \| 0.453 \| 32,599 \|

	### target = production, setting = in_season

	\| model \| covariates \| sWQL \| WQL (pooled) \| MASE \| MdASE \| MAPE \| MdAPE \| cov_80 \| n \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| lightgbm_cov \| yes \| 0.4082 \| 0.1361 \| 1.650 \| 0.593 \| 2.176 \| 0.270 \| 0.751 \| 32,599 \|
	\| lightgbm \| no \| 0.4086 \| 0.1367 \| 1.656 \| 0.595 \| 2.214 \| 0.276 \| 0.749 \| 32,599 \|
	\| chronos2 \| no \| 0.4241 \| 0.1411 \| 1.707 \| 0.630 \| 2.964 \| 0.290 \| 0.777 \| 32,599 \|
	\| chronos2_cov \| yes \| 0.4330 \| 0.1411 \| 1.760 \| 0.687 \| 3.540 \| 0.305 \| 0.755 \| 32,599 \|
	\| naive \| no \| 0.4515 \| 0.1503 \| 1.676 \| 0.583 \| 2.175 \| 0.270 \| 0.693 \| 32,599 \|
	\| ets_lite \| no \| 0.4566 \| 0.1514 \| 1.719 \| 0.686 \| 3.354 \| 0.299 \| 0.625 \| 32,599 \|
	\| mean3 \| no \| 0.4663 \| 0.1591 \| 1.757 \| 0.722 \| 3.377 \| 0.300 \| 0.675 \| 32,599 \|
	\| hist_mean \| no \| 0.5863 \| 0.2106 \| 2.520 \| 1.353 \| 10.545 \| 0.498 \| 0.453 \| 32,599 \|

	### target = yield, setting = ex_ante

	\| model \| covariates \| sWQL \| WQL (pooled) \| MASE \| MdASE \| MAPE \| MdAPE \| cov_80 \| n \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| lightgbm_cov \| yes \| 0.1849 \| 0.1532 \| 2.346 \| 0.767 \| 0.296 \| 0.128 \| 0.743 \| 32,599 \|
	\| lightgbm \| no \| 0.1849 \| 0.1530 \| 2.348 \| 0.764 \| 0.298 \| 0.129 \| 0.750 \| 32,599 \|
	\| chronos2 \| no \| 0.1953 \| 0.1517 \| 2.349 \| 0.767 \| 0.285 \| 0.138 \| 0.789 \| 32,599 \|
	\| chronos2_cov \| yes \| 0.1978 \| 0.1520 \| 2.385 \| 0.821 \| 0.288 \| 0.143 \| 0.762 \| 32,599 \|
	\| ets_lite \| no \| 0.2072 \| 0.1611 \| 2.344 \| 0.780 \| 0.309 \| 0.134 \| 0.600 \| 32,599 \|
	\| mean3 \| no \| 0.2098 \| 0.1661 \| 2.376 \| 0.829 \| 0.307 \| 0.142 \| 0.637 \| 32,599 \|
	\| naive \| no \| 0.2149 \| 0.1690 \| 2.350 \| 0.740 \| 0.310 \| 0.130 \| 0.661 \| 32,599 \|
	\| hist_mean \| no \| 0.2494 \| 0.2174 \| 2.966 \| 1.203 \| 0.339 \| 0.209 \| 0.533 \| 32,599 \|

	### target = yield, setting = in_season

	\| model \| covariates \| sWQL \| WQL (pooled) \| MASE \| MdASE \| MAPE \| MdAPE \| cov_80 \| n \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| lightgbm_cov \| yes \| 0.1829 \| 0.1525 \| 2.330 \| 0.760 \| 0.294 \| 0.127 \| 0.745 \| 32,599 \|
	\| lightgbm \| no \| 0.1849 \| 0.1530 \| 2.348 \| 0.764 \| 0.298 \| 0.129 \| 0.750 \| 32,599 \|
	\| chronos2 \| no \| 0.1953 \| 0.1517 \| 2.349 \| 0.767 \| 0.285 \| 0.138 \| 0.789 \| 32,599 \|
	\| chronos2_cov \| yes \| 0.1973 \| 0.1531 \| 2.385 \| 0.812 \| 0.286 \| 0.142 \| 0.738 \| 32,599 \|
	\| ets_lite \| no \| 0.2072 \| 0.1611 \| 2.344 \| 0.780 \| 0.309 \| 0.134 \| 0.600 \| 32,599 \|
	\| mean3 \| no \| 0.2098 \| 0.1661 \| 2.376 \| 0.829 \| 0.307 \| 0.142 \| 0.637 \| 32,599 \|
	\| naive \| no \| 0.2149 \| 0.1690 \| 2.350 \| 0.740 \| 0.310 \| 0.130 \| 0.661 \| 32,599 \|
	\| hist_mean \| no \| 0.2494 \| 0.2174 \| 2.966 \| 1.203 \| 0.339 \| 0.209 \| 0.533 \| 32,599 \|

	Models: `naive` (last observed value), `mean3` (mean of the last three observed years), `hist_mean` (mean of the
	context), `ets_lite` (damped-trend exponential smoothing), `lightgbm` / `lightgbm_cov` (global LightGBM quantile
	regression on scale-free history features, without / with the setting's covariates, retrained per origin),
	`chronos2` / `chronos2_cov` (`amazon/chronos-2` zero-shot, without / with the covariates).

	## Files

	\| file \| content \|
	\|---\|---\|
	\| `panel/panel.parquet` \| district x crop x season x year area, production, yield (297,505 rows) \|
	\| `series/series.parquet` \| the 12,840 benchmark series, one row per series and year (209,280 rows) \|
	\| `series/covariates.parquet` \| covariates per series-year and setting (418,560 rows) \|
	\| `splits.json` \| origins, context length (23), minimum context (8), windows per origin \|
	\| `forecasts/<model>/<target>_<setting>_<origin>.parquet` \| forecasts of the eight models \|
	\| `leaderboard.csv`, `leaderboard.md` \| the leaderboard above (32 rows) \|
	\| `leaderboard.json` \| the leaderboard plus per-origin (100) and per-crop (1,080) rows and the crop shares of the pooled targets \|
	\| `sources/imd_district_resources.csv` \| the 867 IMD gridded district rainfall resources (one per month, 1951-2023) \|
	\| `evaluate.py`, `requirements.txt` \| evaluation harness \|

	## Schema

	`panel/panel.parquet` (`series/series.parquet` has the same columns plus `series_id`, `ts`, `target`)

	\| column \| type \| description \|
	\|---\|---\|---\|
	\| `state_lgd`, `state_name` \| int16, string \| state (LGD code and name) \|
	\| `district_lgd`, `district_name` \| int32, string \| district (LGD code and name) \|
	\| `crop` \| string \| crop name (56 crops) \|
	\| `season` \| string \| `kharif`, `rabi`, `whole_year`, `summer`, `winter`, `autumn` \|
	\| `year` \| int16 \| agricultural year (July-June) starting in this year \|
	\| `area_ha` \| float64 \| area, hectares \|
	\| `production` \| float64 \| production: tonnes; bales for cotton, jute and mesta; nuts for coconut \|
	\| `yield` \| float64 \| production / area \|
	\| `src` \| string \| `ndap` (NDAP dataset 6820) or `ogd_apy` (data.gov.in resource 35be999b..., filling keys missing from NDAP) \|
	\| `yield_reported` \| float64 \| yield as published in the source \|
	\| `series_id` \| string \| `district_lgd\|crop\|season` (series file only) \|
	\| `ts` \| date \| July 1 of `year` (series file only) \|
	\| `target` \| float64 \| equal to `yield` (series file only) \|

	`series/covariates.parquet`: keys `district_lgd, crop, season, year, setting` (`ex_ante` / `in_season`) and

	\| column \| description \|
	\|---\|---\|
	\| `cov_rain_season` \| log((season rain + 10 mm) / (1951-1996 climatology + 10 mm)); `in_season` only \|
	\| `cov_rain_pre` \| the same for the pre-season months before sowing \|
	\| `cov_rain_prev` \| last year's season rainfall anomaly \|
	\| `cov_price_pre` \| log(median mandi modal price of the crop in the 3 months before sowing / the same months a year earlier); district level, state median as fallback \|
	\| `cov_reservoir` \| national live storage / live capacity of CWC reservoirs on the last report before the cut-off (at most 21 days old) \|
	\| `cov_rain_tier` \| 1 = district IMD gridded rainfall, 2 = state mean of IMD sub-divisions, 0 = none \|
	\| `cov_max_date` \| latest date any covariate value depends on \|
	\| `cutoff` \| the row's information cut-off: the season's sowing month for `ex_ante`, the month after the season's rainfall months for `in_season` \|

	`forecasts/<model>/*.parquet`: `series_id, district_lgd, state_lgd, crop, season, origin, h` (= 1), `ts, q0.1 ...
	q0.9, mean`, `target_value` (the observed value), `mase_scale`, `model`, `target`, `setting` (`univariate`,
	`ex_ante` or `in_season`).

	`sources/imd_district_resources.csv`: `uuid` (data.gov.in resource id), `month`, `api_total` (rows the API reports),
	`records_in_dump` (rows in the snapshot).

	## Coverage

	\| \| \|
	\|---\|---\|
	\| panel rows (district x crop x season x year, with yield) \| 292,333 \|
	\| districts / states / crops \| 671 / 35 / 56 \|
	\| benchmark series (median area >= 50 ha, >= 8 years) \| 12,840 \|
	\| rolling origins (agricultural years) \| 2015, 2016, 2017, 2018, 2019 \|
	\| horizon \| 1 year (next season of the same type) \|
	\| test rows per target \| 32,599 (6,984 / 7,248 / 7,343 / 7,357 / 3,667 for 2015-2019) \|

	Rainfall: district-level (tier 1) rainfall covers 340 districts; the others use state-level sub-division means
	(tier 2). Share of benchmark series-years by tier: {'1': 0.4168, '2': 0.5627, '0': 0.0205}.

	- IMD gridded district rainfall: every monthly file in the snapshot holds only the first 10,000 of about 22,700 API
	rows, so only districts of states sorted first (Andaman & Nicobar .. part of Madhya Pradesh) have district
	rainfall; other districts use the state mean of IMD sub-divisions (to 2017).
	- NDAP indicator columns are unlabeled in the dump; area/production/yield were identified by the identity
	yield = production / area (holds on 100.0% of rows).
	- Agricultural year 2019-20 is partial in the source (about half the districts).
	- PM-KISAN is not a covariate: instalments start in 2019, after the last complete panel year.
	- About 3% of panel values (4-6% of Kharif, Rabi and Whole Year test rows) repeat the previous year's figure exactly.

	## Source data

	Derived from 871 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).

	Providers: Central Water Commission, India Meteorological Department, Ministry of Agriculture and Farmers Welfare, NITI Aayog (NDAP).

	- Source: NITI Aayog, National Data and Analytics Platform (NDAP), dataset 6820: Land Use Statistics : Crop Production Statistics (Ministry of Agriculture and Farmers Welfare). https://ndap.niti.gov.in/dataset/6820
	- Department of Agriculture and Farmers Welfare (Ministry of Agriculture and Farmers Welfare). "District-wise, season-wise crop production statistics from 1997". data.gov.in resource 35be999b-0208-4354-b557-f6ca9a5355de, https://api.data.gov.in/resource/35be999b-0208-4354-b557-f6ca9a5355de. Licensed under the Government Open Data License - India (https://data.gov.in/government-open-data-license-india). Snapshot: May 2026.
	- India Meteorological Department (IMD) (Ministry of Earth Sciences). "Sub Divisional Monthly Rainfall from 1901 to 2017". data.gov.in resource 8e0bd482-4aba-4d99-9cb9-ff124f6f1c2f, https://api.data.gov.in/resource/8e0bd482-4aba-4d99-9cb9-ff124f6f1c2f. Licensed under the Government Open Data License - India (https://data.gov.in/government-open-data-license-india). Snapshot: May 2026.
	- IMD GRID MODEL daily district rainfall, one data.gov.in resource per month (resource list in `sources/imd_district_resources.csv`). Licensed under the Government Open Data License - India.
	- Agmarknet daily mandi prices (data.gov.in resource 35985678-0d79-46b4-9ed6-6f13308a1d24) and CWC reservoir storage (data.gov.in resource 1fc2148c-fc41-46f5-a364-bdc03f77053f). Licensed under the Government Open Data License - India.

	GODL-India does not cover personal information; this dataset holds aggregate statistics only.

	<!-- jm:source -->
	Original data: published by Department of Agriculture and Farmers Welfare, Ministry of Agriculture and Farmers Welfare (district crop production statistics, via data.gov.in and NITI Aayog NDAP); India Meteorological Department, Ministry of Earth Sciences (rainfall); Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare (mandi prices); Central Water Commission, Ministry of Jal Shakti (reservoir storage) 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{judicialmind2026cropbench,
	title = {CropBench: a forecasting benchmark for district crop yield and production in India},
	author = {JudicialMind},
	year = {2026},
	publisher = {Hugging Face},
	howpublished = {\url{https://huggingface.co/datasets/judicialmind/CropBench}}
	}
	```

	<!-- 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 -->