AgrDataSci/tricot-evidence

Tricot evidence

Evidence for plant breeders from on-farm tricot trials: 61 trials, 12 crops, 6 African countries, 18,246 farms and 646,161 farmer rankings (2020-2025), with the farms' climate, fitted variety worths, what traits farmers reward, household surveys around the trials, and video concept-testing surveys. Built by AgrDataSci/tricot-evidence; every decision that changes a value is in its docs/SOURCES.md.

Twenty-five parquet tables, read with DuckDB. Two keys join everything: study_id is a trial, block_id is a farm and starts with its study_id.

import duckdb
duckdb.sql("create secret (type huggingface, provider credential_chain)")  # while the repo is private: uses `hf auth login`
root = "hf://datasets/AgrDataSci/tricot-evidence"   # pin a revision: root + "@<commit>"
T = lambda name: f"'{root}/{name}.parquet'"
duckdb.sql(f"select count(*) from {T('farms')}")

Tables

Table	One row per	Key	Use it for
studies	trial or concept survey (kind)	study_id	crop, country, dates, producer, licence
farms	farmer	block_id	location, planting and harvest dates, gender, age, education, land, market distance, share sold; concept farmers have no location
genotypes	variety in a trial	study_id, genotype	role (check, candidate), release year, origin
rankings	farm × visit × trait × variety	block_id, moment, trait, genotype	the raw evidence; rank 1 to 3, ties allowed
measurements	plot × visit × measured variable	block_id, moment, genotype, variable	what was measured, not ranked: yield, plant counts, harvest and flowering dates, disease scores; names as published, question says what was asked, value_num/value_date are typed
genotype_map, trait_map	a raw name		what each raw name became and why
climate	farm-season	block_id, planting_date	rainfall, dry spells, hot days, tmax, tmin, GDD per growth stage of that season
climate_future	farm-season × scenario × period × quantile	block_id, planting_date	the same features with the season shifted by the CMIP6 monthly deltas
elevation	located farm	block_id	metres
chc_daily	0.05° cell × day	cell_lat, cell_lon, date	daily rain, tmax, tmin 2020 to now for every farm cell
future_climate	0.25° cell × hazard × scenario × period × month	cell_lat, cell_lon	CMIP6 ensemble change vs 1995-2014
varieties	crop × network × trait × variety		Plackett-Luce worth, reliability vs the default check
pairwise	crop × network × trait × variety × versus		chance the first beats the second, interval, farms with both
trait_agreement	crop × network × trait		Kendall tau of the trait's ranking with overall, all farms
farm_trait_agreement	crop × network × trait × farm	block_id	the same tau per farm with its weight; group it any way
tree_nodes, tree_worths, tree_splits, tree_farms	crop × network × trait × set (environment or farmer)	node, block_id	Plackett-Luce trees: where worths change with climate, elevation or farmer; leaf worths; the test for every covariate at every node; which leaf each farm is in
tree_farms_future	farm × scenario × period × quantile	block_id	the leaf the farm falls in when its season is shifted by the CMIP6 deltas, next to its leaf today, margin, its distance to the nearest line, and beyond_range: the shifted season is outside every trial season in the network on beyond_variables, so it has no analog here
concepts, concept_rankings	survey × concept; farmer × concept	block_id, concept	what farmers want that no variety has (maize KE/UG, groundnut and sorghum TZ)
rhomis_households	surveyed household	id_unique	district-level population: land, livestock, market orientation, welfare
rhomis_crops	household × crop	id_unique, crop	what they grow, eat, sell, who controls it

A network is a set of trials connected through shared varieties, so worths are on one scale. Named by country; numbered when a country has more than one.

Standard queries

Farms in a region. Districts come from the admin-boundaries skill. Or a bounding box:

select block_id, crop, latitude, longitude, planting_date, gender, land_ha, sell_share_pct
from farms where country = 'RW' and latitude between -2.2 and -1.7 and longitude between 29.8 and 30.4

What traits decide "best" for a set of farms. The weighted mean of per-farm taus over any subset. Bootstrap the farms for an interval; below about 100 farms say the number is rough.

select t.trait, sum(t.kendall_tau * t.n_effective) / sum(t.n_effective) as tau, count(*) as farms
from farm_trait_agreement t join region_farms using (block_id)
where t.crop = 'maize' group by 1 order by tau desc

Tau 1 means the trait ranking and the overall ranking always agree; 0 means no relation. In Rwanda maize, marketability and grain colour sit near 0.8, maturity near 0.05.

Varieties against the breeder's check.

select variety, p_beats, p_beats_low, p_beats_high, n_farms_both
from pairwise where crop = 'maize' and trait = 'overall' and versus = 'WH 507' order by p_beats desc

p_beats is the modelled chance the variety ranks above the check on a farm. n_farms_both is direct evidence; zero means the comparison runs through other varieties.

Where performance flips. A tree is a prompt, not a target: show the split and ask whether it matches anything the breeder knows. Always show the runners-up from tree_splits.

select node, rule, split_variable, split_at, n_farms from tree_nodes
where crop = 'maize' and network = 'RW' and trait = 'overall' and set = 'environment'

A place with no trial farms. Any point can be placed in a tree without R, because the leaf rules are threshold expressions. It is extrapolation; say so.

1. Daily rain, Tmax and Tmin for the point's 0.05° cell (CHIRPS v3 and CHIRTS-ERA5, the trees' sources), elevation from the Copernicus 90 m DEM.
2. A planting date and season length from the breeder, or the crop's reference length. Stage windows are in config/crops/<crop>.toml, scaled by season length over reference length and rounded.
3. The features, named <variable>_<window>, with the trees' definitions: rainfall_total is the sum of days with at least 1 mm; max_dry_spell the longest run of days under 1 mm; tmax_mean and tmin_mean plain means; hot_days_30 and hot_days_35 days strictly over 30 and 35 °C; gdd the sum of max((Tmax + Tmin)/2 − 10, 0); max_day_rain the wettest day. A window with any missing day is NULL.
4. The leaf: the terminal row of tree_nodes whose rule is true (replace & with and). Report the distance to each threshold on the path; a point within half a degree or 20 mm of a line is on the line. Then the leaf's rows in tree_worths are the trial farms it resembles, on the dimensions the rankings respond to.

For the point's future, shift the daily series by the CCKP monthly deltas for its 0.25° cell (Tmax and Tmin added, rain scaled by percent) and repeat step 3 and 4.

What the trial seasons were like.

select c.block_id, c.planting_date, c.max_dry_spell_reproductive, c.hot_days_30_reproductive,
       c.rainfall_total_season, c.tmin_mean_grain_fill
from climate c join region_farms using (block_id)

The population around the trials. RHoMIS coordinates are 0.1° cells; match on the cell.

select h.* from rhomis_households h
where h.iso = 'RW' and (round(h.cell_lat, 1), round(h.cell_lon, 1)) in
      (select distinct (round(latitude, 1), round(longitude, 1)) from region_farms)

Read before using

• Rankings are farmer judgements of their three plots, not measurements. "Drought tolerance" is what the farmer saw.
• grow_again ("would you plant it next season?") is a second overall judgement, not a trait. It has worths and pairwise odds but no trait agreement row.
• measurements are farmer-recorded and unharmonised: the same quantity has several names across trials (harvestdate/harvestingdate), units are whatever the trial used, and values are not range-checked. Read question and the trial before pooling.
• Trees split where the data splits, at 0.05, among the covariates offered. A split is a place to look, not a zone to breed for; tree_splits shows how close the other covariates came. The farmer trees use only the farms that were asked, often half the trial.
• tree_farms_future is which of today's farms a farm's future season resembles, not a forecast of performance. Means were shifted month by month, so dry-spell features are rough, and a farm within margin of a line is on the line, not across it. A row with beyond_range true is hotter (or rarely colder) than any trial season in the network: the tree still assigns a leaf, but no farm in this data has seen that season. In Rwanda maize that is 8% of farm-seasons by 2040-2059 under SSP2-4.5 and two thirds by 2080-2099 under SSP5-8.5. Analogs stay inside a network, so a Rwandan farm's future is always another Rwandan trial farm, never one in Tanzania.
• 8% of rankings are partial (only the best or the worst named) and are kept as ties.
• Farmers only ranked traits the trial asked about. A trait with no rows was not asked, not unimportant. Use concepts for traits no variety has.
• trait is harmonised across trials; trait_map says what was merged. Local is each trial's own farmer seed and is never the same variety twice.
• A planting date typed after its form was submitted was dropped; planting_date_source says where each date came from. Farms without one have no climate row.
• RHoMIS Rwanda (one project, 2018) has no food security or poverty module and no northern province; its input flags are NULL because they were not asked. Income there is a tenth of neighbours', a question in the source, not a fact.
• Concept surveys cover Kenya, Uganda and Tanzania only. Evidence borrowed from them for another country is labelled as borrowed.

Licence and citation

config/crops/ ships with the tables: the stage windows the climate features and trees used.

The trial tables derive from tricot-data (de Sousa et al. 2025, CC BY-SA 4.0), so this dataset is CC BY-SA 4.0. Every other source allows redistribution with attribution, and each table's provenance is below. Cite the revision you read.

Tables	Source	Licence	Credit
studies, farms, genotypes, rankings, measurements, maps, fits, trees	tricot-data	CC BY-SA 4.0	de Sousa et al. (2025)
concepts, concept_rankings	CIMMYT Dataverse deposits (maize, groundnut, sorghum)	CC BY 4.0	Rutsaert et al., CIMMYT
rhomis_households, rhomis_crops	RHoMIS v2.0	CC0 1.0	Gorman et al.
chc_daily, climate, climate_future	CHIRPS v3.0 (Funk et al. 2026, data) via dynamical.org; CHIRTS-ERA5 via aaguilar90/chirts-era5-daily	CC BY 4.0	"UCSB Climate Hazards Center CHIRPS version 3.0 data processed by dynamical.org"; Climate Hazards Center, UC Santa Barbara (2024), CHIRTS-ERA5, experimental
elevation	Copernicus DEM GLO-90	ESA Copernicus DEM licence	produced using Copernicus WorldDEM-90 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved
future_climate, tree_farms_future	World Bank CCKP CMIP6 0.25°	CC BY 4.0	World Bank Climate Change Knowledge Portal; CMIP6 model output of the WCRP Working Group on Coupled Modelling (CC BY 4.0 since October 2022)

RHoMIS rows are households at 0.1° with income and food security, exactly as the source publishes them under CC0; nothing here adds to that exposure.