Configuration Reference

Reference for every YAML key consumed by the three acestor pipelines. Each field lists type, default, and why it exists. Source of truth is the configs.py module for each pipeline — this document is derived from those dataclasses.

Three pipelines are documented:

1. dengue_prep — builds prepared_data/ from raw case + weather sources.

2. dengue — trains models, computes thresholds, renders maps and the HTML brief.

3. dengue_downscale — takes parent-level predictions from a dengue run and splits them onto child regions.

Each pipeline has a pipeline.name matching one of the above; the loader dispatches on it.

Common top-level keys (state, pipeline, run, logging, storages, email) are shared and documented once at the end.

---

1. dengue_prep

Config sample: configs/ka_district_prep.yaml, configs/od_district_prep.yaml. Dataclasses: pipelines/dengue_prep/configs.py.

data.prepared_data

Where prep writes its outputs. Consumed by the downstream dengue pipeline via the same block.

key	type	default	why
base_dir	str	prepared_data	Root folder for the per-region-type subfolders (district/, zone/…).

data.case_download

Fetches raw line-list files from a case source.

key	type	default	why
enabled	bool	false	Master switch. When false, prep assumes files are already staged under source_path.
source_mode	str	""	Selects a case-source plugin: dashboard, filesystem, a dotted module path, or a direct file path. Empty → legacy backend.
source_backend	str	filesystem	Legacy path used only when source_mode is blank; filesystem or s3.
source_path	str	"" (falls back to DENGUE_PREP_CASE_SOURCE env)	Root path or URL the source plugin reads from.
base_url	str	plugin-specific	For dashboard mode: dashboard API root, e.g. https://apps.artpark.ai/disease-dashboard.
disease	str	Dengue	Dashboard disease slug.
date_start	str	""	Dashboard: inclusive start of the fetch window (YYYY-MM-DD).
date_end	str	"" (→ run_date)	Dashboard: inclusive end. Blank tracks the current run.
chunk_days	int	plugin default	Dashboard pagination window; smaller values ease per-request payload.
backfill_days	int	plugin default	Dashboard: how many days before date_start to re-fetch to catch late-arriving reports.
selected_region_id	str	""	Restrict the dashboard fetch to a single region (debugging).
cache_enabled	bool	true	Turn off to force re-download on every run.
cache_dir	str	./cache/raw_case	Local cache root.
cache_strategy	str	local_first	local_first skips the network when the cache hits.
filesystem_base_path	str	""	Absolute or relative filesystem root when source_backend=filesystem.
s3_bucket, s3_prefix	str	""	S3 backend location.
source_paths	list[str]	[]	Explicit list of files a plugin should ingest.
dest_relpath	str	datasets/raw_linelist_data/linelist	Relative destination under the run artifact root.

data.case_parse

Parses raw case files into cases_daily.csv.

key	type	default	why
region_types	list[str]	required	Non-empty list of region tiers to emit; each becomes a subfolder under prepared_data/.
date_start	str	""	Inclusive start filter on the case date. Empty → no lower bound.
date_end	str	""	Inclusive end filter. Empty → no upper bound.
date_column	str | list[str]	["Sample Collected Date"]	Candidate column(s) for the case date. Parser tries in order; first one that parses cleanly wins. String is accepted for back-compat.
region_id_column	str	""	When set and present in the file, this column is trusted verbatim as the resolved region_id — skips LGD lookup, geocode, spatial join.
lgd_code_column	str	""	Column with an LGD code. Present → used to resolve to region_id.
lat_column	str	Latitude	Fallback spatial-join latitude column.
lon_column	str	Longitude	Fallback spatial-join longitude column.
header_row	int	0	0-indexed row containing headers; use 1 for banner-row IHIP exports.
filters	list[dict]	[]	Row filters: each entry {column, values}; AND across entries, OR within values.
geocoding	dict	disabled	Google-geocode + spatial-join resolver; see below.

data.case_parse.geocoding

Runs only when a row has neither LGD code nor region_id_column and only free-text address fields are usable.

key	type	default	why
enabled	bool	false	Opt in explicitly.
address_fields	list[str]	required when enabled	Columns concatenated to build the primary query.
fallback_address_fields	list[str]	[]	Second-pass columns when the first fails validation.
cache_file	str	./cache/geocode_cache.json	On-disk memoisation of Google responses.
extra_stopwords	list[str]	[]	Extra tokens stripped before validation.
require_address	bool	true	Drop rows whose first address field is empty.
require_validation	bool	true	Drop rows whose Google response fails token validation.
bounds	list[4]	null	[min_lat, min_lon, max_lat, max_lon] viewport bias for the geocoder.
restrict_admin_area_tokens	list[str]	[]	Substrings that must appear in the formatted address; hard reject.

data.weather_download

Fetches gridded weather. Two source families: CDS (reanalysis-era5-land) and openmeteo.

key	type	default	why
enabled	bool	false	Master switch.
source_mode	str	filesystem	Plugin selector: filesystem, cds, openmeteo.
source_backend	str	filesystem	Backend for the filesystem plugin: filesystem or s3.
temperature_unit	str	celsius	Unit the source returns (celsius/kelvin). The parser normalises on-disk to Kelvin so downstream code sees one convention.
precipitation_unit	str	mm	Unit the source returns (mm/m). Normalised to metres on disk.
netcdf_cache_path	str	""	Local NetCDF cache path; empty → plugin default.
parsed_output_path	str	""	Where per-region CSVs land after parsing.
cds_variables	list[str]	env GBA_CDS_VARIABLES	CDS variable names.
region_bounds	list[4]|null	null	[N, W, S, E]; null → auto-compute from the geojson.
region_type	str	district	Subfolder under the geojson root used to compute bounds.
w_params	list[str]	[t2m, d2m, tp]	Short names of NetCDF variables the parser extracts.
threshold_km	float	25.0	Max distance from a region’s boundary a grid point may lie to still be attributed.
bounds_resolution_deg	float	0.1	Snap auto-computed bounds to this grid (0.25 for ERA5, 0.1 for ERA5-Land).
start_date	str	2015-01-01	Inclusive fetch start.
end_date	str	""	Inclusive end. Empty → run_date.
cache_enabled	bool	true	Turn off to force re-download.
cache_dir	str	./cache/raw_weather	Local cache root.
cache_strategy	str	local_first	local_first skips the network on cache hit.
source_path, filesystem_base_path, s3_bucket, s3_prefix, source_storage, source_prefix, source_paths, dest_relpath	str/list	"" / datasets/raw_weather_data	Location knobs mirroring case_download.

data.weather_parse

Daily aggregation of downloaded weather.

key	type	default	why
region_type	str	district	Region tier the daily rows are keyed to.
weather_variables	list[str]	["2mTemperature","totalPrecipitation","2mDewpointTemperature"]	Variables to aggregate.
daily_agg	list[dict]	mean for temps, sum for precip	Per-variable daily aggregation ops.

data.geojson

key	type	default	why
base_path	str	—	Root of the geojson layer; subfolders per region type.

data.date_range.start

key	type	default	why
start	str	—	Global anchor for the prep window; passed to plugins that need a fetch start when date_start is blank.

---

2. dengue

Config sample: configs/ka_district.yaml, configs/gba_zone.yaml. Dataclasses: pipelines/dengue/configs.py.

data.prepared_data

key	type	default	why
base_dir	str	prepared_data	Root of the prep output the pipeline reads.
region_type	str	district	Subfolder to load from (district, zone, corp, ward, subdistrict).

data.case_parse

key	type	default	why
region_types	list[str]	required	Non-empty, no duplicates. Order matters: the last entry drives cutoffs and the sampled CSV that maps/thresholds consume.
date_start	str	required	Inclusive start.
date_end	str	required	Inclusive end; empty string → parser uses today.

data.case_sufficiency

Early abort when the parsed case data is too thin to model.

key	type	default	why
enabled	bool	true	Turn off to run the pipeline on tiny/synthetic inputs.
min_total_rows	int	30	Row-count floor.
min_distinct_regions	int	2	Region-count floor.
min_date_span_days	int	14	Temporal span floor.
case_column	str	case	Which column carries the case count.
region_column	str	""	Empty → derived from region_type.
date_column	str	""	Empty → derived from region_type.
max_staleness_days	int	0	Max acceptable gap between run_date and the freshest observed case row. Catches run_date running ahead of ingest. 0 disables.

cutoff

key	type	default	why
case_min_regions	int	2	Minimum regions with observations before the case cutoff is honoured.
weather_min_regions	int	5	Same, for weather.

thresholds

key	type	default	why
region_type	str	zone	Region tier the thresholds are computed for.
methods	list[str]	[historical, prev_nweeks]	Which methods to compute; each writes its own threshold rows. Registered names: historical, prev_nweeks, weighted_baseline.
classification_method	str	who	Zone-labelling scheme: who, icmr, or percentile. Unknown values raise (previously silently produced WHO output — issue #62).
percentile_cutoffs	list[float]	[25.0, 50.0, 75.0]	For classification_method: percentile only. N cutoffs → N+1 bands, computed per region against its own weekly history. Common override: [50, 75, 90].
n_weeks	int	4	Prev-N-weeks window.
historical_n_years	int | null	4	Historical lookback in years. null → all history.
excluded_years	list[int]	[2020, 2021]	Years dropped from the historical baseline (COVID gaps).
included_years	list[int]	[]	Empty → no restriction; non-empty → whitelist.
list_alpha	list[float]	[1.0, 2.0]	Sigma multipliers for the WHO/ICMR banding.
recent_weeks	int	4	weighted_baseline recent-window length.
sd_window_weeks	int	8	weighted_baseline std-dev window.
weight_recent	float	0.7	weighted_baseline recent-mean weight.
weight_seasonal	float	0.3	weighted_baseline seasonal weight (52-week-lag mean).
method_configs.<name>	dict	inherit	Per-method overrides for n_weeks, historical_n_years, excluded_years, included_years, recent_weeks, sd_window_weeks, weight_recent, weight_seasonal. region_type, methods, classification_method always inherit from the base block.

model

key	type	default	why
spatial_res	str	zone	Region tier the models train and predict on.
models	list[str]	["tse"]	Subset of nbr, tse, rf, xgb, timesfm. nbr was retired from the default set (issue #63) but is still opt-in via models: [nbr, ...].
ensemble	str	mean	Registered ensemble strategy or none. none is incompatible with output=ensemble.
output	str	ensemble	ensemble, per_model, or both.
tune	bool | str	false	true → always Optuna-retune. false → use cache when fingerprint matches, otherwise retune. "never" → trust cache regardless (hindcast / production reuse); fails loud on cache miss.
n_trials	int	100	Optuna trials when tuning runs.
debug	bool	false	Emit per-model intermediate CSVs under artifacts/debug/<model>/.
data_features	list[str]	[]	Columns fed to the tabular models.
years_to_exclude	list[int]	[2020, 2021]	Training years dropped.
years_to_include	list[int]	[]	Empty → no restriction; non-empty → whitelist.
lag.lag_temp	list[int]	[12]	Temperature lag weeks fed as features.
lag.lag_rainfall	list[int]	[4]	Precipitation lags.
lag.lag_humidity	list[int]	[4]	Dewpoint lags.
lag.lag_cases	list[int]	[]	Case-count lags (autoregressive).
clip_multiplier	float | null	null	Upper clip in recursive forecast: cap each step at clip_multiplier * max(train cases). null = floor at 0 only (default; reference parity).
freeze_weather_at_origin	bool	true	Persistence: freeze weather/exogenous lag features at the forecast origin for every future week. Matches upstream vbd-modelbench. Set false only for hindcasts that deliberately leak future weather.

model_configs.<model>

Per-model overrides. Keys mirror the model block: data_features, years_to_exclude, years_to_include, tune, n_trials, debug, clip_multiplier, freeze_weather_at_origin, and the lag.* sub-keys. Unspecified keys inherit from model. Any key not listed in model.models is rejected at build time (fail-fast; the error message names both the YAML fix and the --set CLI trap that re-creates a removed key).

model_configs.timesfm

TimesFM 2.5 foundation model settings. All optional — omit the whole block and defaults apply.

key	type	default	why
huggingface_repo_id	str	google/timesfm-2.5-200m-pytorch	HF repo the checkpoint is pulled from.
revision	str	env TIMESFM_REVISION or pinned SHA 1d952420fba87f3c6dee4f240de0f1a0fbc790e3	Must be a full 40-hex commit SHA. Tags/branches are rejected — checkpoints can move under a floating ref.
cache_dir	str	.cache/timesfm	Project-local, gitignored.
max_context	int	1024	Truncate each region’s series to this many trailing weeks before inference.
per_core_batch_size	int	32	Torch batch size inside the worker.
min_context_weeks	int	52	Regions with fewer weeks of history are skipped (all-zero regions still get zero predictions).
timeout_s	float	300	Subprocess wall-clock timeout.
max_regions_per_batch	int	128	Cap on regions per worker call — bounds worker memory.

assess

key	type	default	why
total_<region>_regions	int	—	Denominator for assess ratios. Keys recognised: total_corp_regions, total_zone_regions, total_ward_regions, total_district_regions, total_subdistrict_regions.

maps

key	type	default	why
enabled	bool	true	Set false to skip static PNG generation. The HTML report renders interactive D3 maps from embedded GeoJSON regardless; PNGs remain useful for PDF/offline distribution.
output_dir	str	plots	Where PNG choropleths land.
figure_title	str	Dengue risk map	Suptitle line 1; line 2 is the prediction date.

report

key	type	default	why
document_title	str	<pipeline.title> — summary report	Report <title>. Falls back to a generic string when both this and pipeline.title are empty.
primary	str	ensemble	Which prediction feeds maps and the report. Accepts ensemble or a model key. Short aliases (timesfm, ensemble) resolve to the internal model name.
threshold_method_for_report	str	historical	Threshold method the report tables and headline bands use.

report_distribution

Free-form dict rendered into the report header (system_name, organization, state, region, department, contact_email, footer_note). All fields are strings; empty strings are omitted.

---

3. dengue_downscale

Config sample: configs/ka_district_to_subdistrict.yaml. Dataclasses: pipelines/dengue_downscale/configs.py.

run

key	type	default	why
source_run_id	str	—	Run ID whose parent-level predictions are downscaled. Sentinel "latest" picks the newest artifact matching parent_level and the finer brief tolerance (fixed in commit 6d96610).

downscale

key	type	default	why
parent_level	str	required	Singular region tier the source predictions live at (e.g. district).
child_level	str	required	Singular child tier to split onto (e.g. subdistrict). Must differ from parent_level.
window_weeks	int	4	Look-back window (weeks of child-level cases) used to derive per-child shares of parent cases.
historical_fallback_weeks	int | null	null	Longer window tried when the primary window has zero cases across all children of a parent (issue #86). Must strictly exceed window_weeks. null disables.
cases_csv	str	prepared_data/mandal/cases_daily.csv	Child-level case history used to derive shares.
geojson_base_path	str	ap_datasets/geojsons/geojsons_AP	Root of the geojson layer used for the child map.
on_missing_parents	str	error	error or warn. Controls behaviour when a parent has no children in the geojson.

thresholds

Same schema as the dengue pipeline’s thresholds block, applied to child-level predictions. method_configs.<name> overrides are supported identically.

downscale_maps

key	type	default	why
enabled	bool	true	Set false to skip child-level PNG choropleths (interactive maps still render).
output_dir	str	outputs/maps	Where PNGs land.
figure_title	str	Dengue risk map	Suptitle.

---

Common top-level keys

state

Two-letter state slug used in artifact paths and report copy. Free-form.

pipeline

key	type	default	why
name	str	—	Selects the pipeline: dengue, dengue_prep, or dengue_downscale.
title	str	""	Human title used in the report header when report.document_title is blank.
display_name	str	""	Alternate hint used the same way as title.

run

key	type	default	why
run_date	str	"" → today	Anchor date for the pipeline. Blank tracks wall clock.

logging

key	type	default	why
level	str	INFO	Python logging level (DEBUG, INFO, WARNING, ERROR).

storages.artifacts

Where the pipeline writes per-run artifacts.

key	type	default	why
kind	str	—	filesystem or s3.
filesystem.base_path	str	—	Root path when kind=filesystem. Each run gets an isolated subfolder keyed by --run-id.
s3.bucket, s3.prefix	str	—	Bucket / key prefix when kind=s3.

email

key	type	default	why
enabled	bool	false	Turn on to have the report step send the rendered HTML via SMTP. Other keys (smtp_host, smtp_port, to, from, subject, credentials via env) are consumed by the email step directly.