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.
187-grounding
Detection-format defect localization on metal part — 1346 items (1064 good + 282 defective),
derived deterministically from the binary segmentation masks of AI4Manufacturing/187.
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 one of scratches, anomalous, parts mismatch, total rust, bend and parts mismatch, major rust, defective painting, hole. Good images have annot = [] (1064). 92 defective images are
multi-instance (≥2 boxes).
Verified: every box list re-derived independently from the mask at build — byte-identical on all
1346 rows; goods all []; every box within image bounds.
Blind baseline (report alongside any score). 79% of items are defect-free, so a blind
"always []" guesser scores 79% on the presence decision — yet every one of the 282
defective images scores 0 without correct boxes (exact-match graded). The mostly-good prior is
inherent to the source (industrial QC).
Resolution / legibility. Train at the native resolution the images ship in; 9% of gold
boxes have a min side < 16 px at native (0% < 8 px) — the smallest scratch/paint fragments —
and downscaling below native collapses them.
| field | meaning |
|---|---|
query |
task prompt + JSON output spec (closed class list) |
image |
the metal part 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, r187_record_id, defect_type, n_instances |
Provenance
Built deterministically (no LLM/teacher; reasoning is null) from AI4Manufacturing/187
— MPDD metal-parts surface inspection — 1,346 photos over 6 part categories, 8 defect types + good; each anomalous image has a paired binary pixel segmentation mask (binarized at gray>127 (the source masks are clean 0/255); the defect TYPE comes from metadata (one type per image, raw underscores normalized to spaces — raw label kept in metadata.defect_type_raw)).
Generator: annotate/187/build_187_derived.py in forge_model; machine gates
annotate/187/verify_187.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 1024×1024; 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
187).
- Downloads last month
- 14