[masonearles]

Follow-up to the "why is the count still 731" question. The site only shows what is committed in hf_datasets.json, and comparing that file against the Project-AgML Hugging Face org (508 repos) turned up drift in both directions. Two commits, reviewable separately:

1. Six hf_link values pointed at repos that don't exist (b156cdd)

The site name and the Hub name differ by a typo or doubled suffix, so the Hugging Face button on these cards 404'd. Only hf_link changes; names, sample images and performance files stay as they are.

site record	linked (missing)	now links
teaLeafBD_disease_classification_classification	…_classification_classification	teaLeafBD_disease_classification
mango_growth_detection	mango_growth_object_detection	mango_growth_detection
banana_bunch_maturity_classification_classification	…_classification_classification	banana_bunch_maturity_classification
growliflower_cauliflower_segmentation	growliflower_cauliflower_segmentation	growliflower_caluiflower_segmentation (Hub typo)
vegann_multicrop_presence_segmentation	vegann_multicrop_presence_segmentation	vegann_mulitcrop_presence_segmentation (Hub typo)
OpenEar_Projection_segmentation	OpenEar_Projection_segmentation	OpenEar_projection_segmentation (Hub casing)

Renaming the three misspelled Hub repos (the Hub redirects old names) would be the cleaner long-term fix; I didn't touch the org.

2. Eighteen Hub datasets had no site record at all (1375459, 9ffc9bb)

Pushed between June and October 2026 and never added to hf_datasets.json. Records are built from each dataset's own Hub card, the AgML library's public_datasources.json (leaf_counting_denmark) and the cited paper or Mendeley page; anything the sources don't state is left blank rather than guessed.

Vision (11): ucm_land_use_classification, siri_whu_scene_classification, AvoAir_DB_unlabeled, AvoAir_DB_annotated, tomato_disease_detection, YOLOPOD_Soyabean_Pod_Counting_Dataset, MTDC_Maize_Tassels_Detection_Counting_Dataset, CBDA_Cotton_Ball_Detection_Augmented, leaf_counting_denmark, fruits_detection_australia, sorghum_lidar_point_cloud_segmentation.

Vision-language (7), same schema as the existing VLM records: AgriBench, TomaMMU, PlantMicro, MIRAGE, LeafNet, AgroOmni, PlantExpertVQA. VLM datasets on the site go from 6 to 13.

Things worth a second look:

• fruits_detection_australia is the ACFR Orchard Fruit Dataset (Bargoti & Underwood): classes (apple, mango, almond) and exact image count (1120+1964+620 = 3704) match. The third commit fills location (Warburton VIC / Bundaberg QLD / Mildura VIC), cameras and platform from the ACFR page and the paper. The ACFR page states no license, only "please cite", so license stays blank.
• sorghum_lidar_point_cloud_segmentation is identified as Patel, Cho & Park's Mendeley dataset (10.17632/pfnfzrmrg7.1); the Hub card is bare. No point_cloud_sample_url, so the 3D viewer won't show for it.
• AvoAir_DB_annotated: its Hub card says the boxes are "avocado fruit size variations", but the Avo-AirDB paper annotates tree crowns (large/medium/small), so I used crop_detection, not fruit detection.
• AgroOmni has qa_type: [] and task_dimensions: [] because its card lists counts (14 task types, 56 categories) but not the names or answer mechanisms.
• PlantMicro uses sensor_modality: "microscopy", a new value in the Modality filter; it is light/fluorescence/electron microscopy, not RGB photography.
• Environment for the sorghum (greenhouse LiDAR) and tomato (greenhouse) datasets is left as ""/field pending the greenhouse-vs-field convention flagged in AgML-Journal-Ingestion Feature/new datasets #10.
• Sample grids (third commit) for the ten vision datasets except the sorghum point cloud, generated with the ingestion repo's renderer from Hub parquet streaming / the COCO annotations. The seven VLM datasets have none, like the existing VLM records.

Not registered, both duplicate uploads of datasets the site already lists: potato_leaf_blight (same Laizer et al. Irish potato data as potato_leaf_blight_classification; 55,453 rows vs the complete 58,709) and date_fruit_detection (1,071 images, all cultivar "bouisthami" from sites S1-S4, 4 maturity classes: the Bouisthami subset of Zarouit et al.'s Zenodo dataset that date_fruit_maturity_detection already carries in full with 9,010 images). Both should be deleted from the org; my token lacks delete rights.

Separately, the AgML python library's public_datasources.json lists 9 names that don't appear verbatim in the registry. Seven of them are here under other names: the four gemini_*_2022 datasets are the site's gemini_flower/leaf/pod/plant_detection (same image counts and docs URL, just no _2022 suffix); onion_leaf_classification / chilli_leaf_classification are the augmented splits of the same Aishwarya & Reddy dataset the site lists as COLD_onion/chili_leaf_disease_classification (originals, 815 / 532 images); papaya_leaf_disease_classification is the classification view of BDPapayaLeaf, listed here as papaya_leaf_disease_detection (same five classes). Only autonomous_greenhouse_regression (4TU lettuce challenge, 389 images) and susnato_plant_disease_detection_processed (a third-party HF mirror) are genuinely absent from both the Hub and the site. Not touched here.

Verified locally

• npm run generate-datasets, npm run generate-embeddings (6668 records), npm run test:unit 23/23, npm run build success. Built hf_datasets.json has 747 records (was 729).
• python src/audit_hf_registry.py from AgML-Journal-Ingestion (new, see that repo's PR) reports 0 broken links and only the two deliberately skipped repos as unregistered.

🤖 Generated with Claude Code