+ Select models and countries. Real scores need the full FTW benchmark tree
+ on the API host (BENCHMARK__DATA_ROOT): chips GeoParquet,
+ data_config (STAC URLs), and label_masks/instance/ per country.
+
+ With BENCHMARK__AUTO_DOWNLOAD=true, the API looks for
+ chips_{country}.parquet on Source Cooperative (not the older
+ boundaries_*-only drops). You still need STAC URLs and label masks for a full
+ run. Use BENCHMARK__ALLOW_DEMO=true for placeholder scores without data. Runs
+ are sequential per chip (download + inference + scoring); tens of seconds
+ to a few minutes per chip is normal. Matching uses a default IoU of
+ 0.25 because auto-picked Sentinel scenes are not the benchmark’s original
+ imagery (overlaps are often <0.5).
+
split column. This choice keeps only
+ rows in that partition, then max chips is applied.
+ Test — held-out set (usual benchmark).
+ Validation — validation fold.
+ Train — training chips (not unbiased
+ generalization). If the file has no split column, this setting has no effect.
+
+ Gold outline = chip footprint; green = ground truth; blue = model predictions. Requires
+ a deployed API that returns map_geojson and at least one scored chip.
+
+ No map layers in this response. If Chips is above zero, open
+ Raw JSON and confirm a top-level
+ map_geojson object. If it is missing, restart the API after updating to a
+ build that includes map support. Also check Include map is on,
+ max_chips ≤ 40, and see the “Map (GeoJSON)” note above if any.
+
{{ JSON.stringify(result, null, 2) }}
+