[masonearles]

Why

The home page charts count maize (42) and corn (17) as separate crops, and draw a 70-character bar label, crop image collection (unlabeled; for pretraining and foundation models), in the task chart. Both come from unconstrained metadata in hf_datasets.json. The pipeline side was fixed in Project-AgML/AgML-Journal-Ingestion#8 (merged); this PR applies the same canonicalisation to the registry already on the site.

What changed

static/data/hf_datasets.json — run through src/registry_vocabulary.py from the merged ingestion main. 148 of 725 datasets touched, only the crop_types and agricultural_task fields.

crop	before	after
maize	42	57
rice	42	43
soybean	34	36
grape	23	35
tomato	33	34
distinct crop strings	672	618

Crop labels merged away, by dataset count: corn 17, grapevine 9, canola 6 (→ rapeseed), paddy 3 (→ rice), brassica napus 3, then grapes, brinjal, urad bean, capsicum, soybeans and a dozen misspellings at 1–2 each. 86 further datasets only had capitalised Latin names lowercased (the site already lowercases for display, so nothing visible changes for those).

agricultural_task	before	after	n
crop image collection (unlabeled; for pretraining and foundation models)	→	""	18
plant part segmentation	→	plant_organ_segmentation	1
fruit_maturity_detection	→	maturity_detection	1
crop and weed segmentation	→	crop_and_weed_segmentation	1

The 18 image_safari entries now match the other 41 image_safari entries, which already carry "".

Derived files, regenerated exactly as npm run build's prebuild does:

• static/data/performance/global.json — crop_types copies for the leaderboard filter.
• static/data/embeddings/{meta.json,vectors.bin} — semantic search index. Record count goes 6286 → 6647 because the committed snapshot predated the October data batches; this is the first regeneration since 2026-09-25.

datasets.json and performance/index.json regenerated byte-identical.

Verification

• npm run test:unit: 23/23 pass
• npm run build: success; built output contains zero occurrences of the long label and zero corn crop entries
• Diff of hf_datasets.json contains only crop_types / agricultural_task lines

Side effects

• Bookmarked dataset-page links filtering on corn will match nothing; new links use maize.
• Searching "corn" still works: semantic search embeds dataset names and descriptions, not just crop labels.

🤖 Generated with Claude Code