You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
185-grounding
Detection-format defect localization on production part surface — 3335 items (2979 good + 356 defective),
derived deterministically from the binary segmentation masks of AI4Manufacturing/185.
The model outputs boxes as text; defect-free images must output [] — detection rejection is
part of the task.
Task
"Locate every defect." annot is a JSON list of {"type": ..., "bbox_xywh": [x, y, w, h]} in the
image's pixel coordinates (origin top-left; see metadata.image_wh), one entry per defect instance
(connected components after proximity grouping; sub-15-px groups denoised), sorted (type, x, y); the
type is a single class defect. Good images have annot = [] (2979). 30 defective images are
multi-instance (≥2 boxes).
Verified: every box list re-derived independently from the mask at build — byte-identical on all
3335 rows; goods all []; every box within image bounds.
| field | meaning |
|---|---|
query |
task prompt + JSON output spec (no type list (binary)) |
image |
the production part surface photo (no overlays) |
annot |
JSON box list, [] when defect-free |
reasoning |
null — deterministic |
cate / task |
B / T-B2 |
metadata |
source, image_sha256, image_path, image_wh, r185_record_id, defect_type, n_instances |
Provenance
Built deterministically (no LLM/teacher; reasoning is null) from AI4Manufacturing/185
— KolektorSDD2 production-part surface inspection — 3,335 photos, binary good vs anomalous (surface cracks/defects); each anomalous image has a paired binary pixel segmentation mask (binarized at gray>40).
Generator: annotate/185/build_185_derived.py in forge_model; machine gates
annotate/185/verify_185.py (all green at build: boxes/gold re-derived byte-identical, clean
regions contain zero defect pixels, mcq gold = the true panel, no leaked-artifact vocabulary, image
set sha256-disjoint from all other cached AI4Manufacturing datasets).
Resolution. Images are native (~230×640; not downscaled); all coordinates live in that pixel space (see metadata.image_wh).
Query diversity. query is drawn from a fixed pool of surface variants (paraphrases preserving
the task + answer format), selected by an independent per-record hash; a machine gate confirms no
template correlates with the gold.
The repository name is an internal task code (the source dataset's code is
185).
- Downloads last month
- 4