FaroukMoc2/jev-stage2-image-beans-pilot
Beans: one natural question per image Open the corrected preview. natural_v4 is the recommended and default preview: 100 original images, 100 rows, one three-way condition-class Choice question per image. All targets come directly from the source labels column (34 angular leaf spot, 33 bean rust, 33 healthy). Original image bytes and source annotations are unchanged. Example question: “Which source-defined condition class describes the bean leaf?” Options: angular_leaf_spot… See the full description on the dataset page: https://huggingface.co/datasets/FaroukMoc2/jev-stage2-image-beans-pilot.
Beans: one natural question per image
natural_v4 is the recommended and default preview: 100 original images, 100 rows, one three-way condition-class Choice question per image. All targets come directly from the source labels column (34 angular leaf spot, 33 bean rust, 33 healthy). Original image bytes and source annotations are unchanged.
Example question: “Which source-defined condition class describes the bean leaf?” Options: angularleafspot, bean_rust, healthy. The answer is stored separately in targets.
There are no Boolean restatements of the class question, artificial image-group count scores, or forced 13-question bundles. Beans does not need all three primitives. Across the broader vision corpus, multiple questions should cover distinct useful facts in naturally richer documents, charts, interfaces, game states and annotated scenes. Primitive coverage is monitored across the corpus; source evidence takes precedence over shape quotas.
The earlier historical_v1, balanced_v2 and varied_v3 configs are retired experimental artifacts. Their repetitive question multiplication and numeric balance results are not recommended designs. They are retained for traceability only.
Validation: all 100 source targets and image identities checked; parquet roundtrip preserves full canonical records and original bytes. All 300 native LLaVA candidate encodings pass at 692–695 tokens, with one image each and no truncation. Complete policy and audit receipts are in natural-v4/metadata/.
This remains a smoke-only browsing preview, with zero approved training rows. Original visual-review flags and uncertainty are retained. Causal source grouping, cross-split overlap, expert label adjudication, production release and model calibration evaluation remain pending. No severity ratings or model-generated ground truth are used.
Source: AI-Lab-Makerere/beans revision 27aa014ce09b193e1a6f58112d4a66e0eddb69c5. Publisher MIT evidence and license are preserved. For correction/removal, open a dataset discussion with the row ID or original-image hash.