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.
D15-mcq
Mask-grounded multiple choice (Set-of-Mark style) for industrial defect localization — 2,828 items, derived
deterministically (no LLM/teacher involved) from the semantic masks of
AI4Manufacturing/D15 (DefectSpectrum). Exact-match
gradable → directly usable for SFT and RLVR-style training.
The repository name is an internal task code. See Provenance below.
Query diversity (2026-07-11). The
queryfield is drawn from a pool of 19 surface variants for this task (paraphrases that preserve the task and answer-format exactly; the answer-format directive is held verbatim), each selected by an independent per-record hash. This replaces the earlier 4-template design to prevent instruction-format overfitting; answers, images, ids, and all provenance are unchanged. A machine gate inverify_*.pychecks that no template correlates with the gold (binomial z < 4.5).
Task
One item per (defective D15 record, defect type). The image is a 2×2 grid of views A–D of the same
product photo; each view overlays one red candidate region mask (SoM style: translucent fill + outline +
corner letter). Exactly one view overlays the true mask for the named defect type — in both location and
extent (the query says so explicitly). annot is the correct letter.
The three negatives per item are hard by construction, drawn in preference order from:
| tag | construction |
|---|---|
othertype |
the true mask of a different defect type in the same image, presented as a candidate for this type (hardest) |
shift |
true mask translated by ~0.7–1.6× its bbox extents (border-clipped, ≥60% of pixels kept) |
fliplr / flipud / rot180 |
true mask mirrored across the image axes |
dilate |
true mask grossly over-grown (bounded at 3% of the image dimension; must be ≥2.5× the true area) — a right-location, wrong-extent decoy |
erode |
true mask shrunk to <60% of its area |
Every negative is guaranteed wrong (IoU vs. truth < 0.35, except dilate which is wrong by extent) and panels are
mutually distinct (pairwise IoU < 0.7). Gold letters are balanced via id-hash: A 681 / B 669 / C 740 / D 738. Every independent choice —
query surface variant, the other-type decoy draw, and the assignment of negatives to slots — is
drawn from its own salted hash, so neither the query text nor the visible arrangement of negative
kinds correlates with the gold position (verified: template×letter at chance; negatives appear in
canonical construction order at the 1/6 shuffled expectation, vs 1.0 in v1/v2; worst single
kind-at-letter rule at the 1/3 conditional chance). Both the gold and every negative overlay must
render ≥30 visible pixels outside the letter label after downscale. Items whose mask covers
35% of the frame, whose gold overlay would be smaller than
30 rendered pixels after panel downscale (unanswerable), or where 3 sound negatives could not be built were **skipped (255 type-instances)** rather than shipped degenerate.
DS-DAGM panels use a histogram-equalized base image (identically in all four views): several DAGM defect
classes are near-invisible in the raw render, which would otherwise make the item unanswerable.
metadata.equalized_base records this.
Records
2,828 items (single train split): DS-MVTec 1,422 · DS-VISION 1,097 · DS-DAGM 268 · DS-Cotton-Fabric 41.
Revision notes (2026-07-09): two adversarial review rounds. v1→v2: query template shared the gold-position hash (query text revealed the answer) + 18 invisibly-small golds. v2→v3: negatives filled slots in fixed preference order (the visible arrangement of negative kinds decoded the gold with 100% accuracy), the other-type decoy draw shared the gold-position seed (parity leak on 3-type records), and the letter label could paint over a tiny gold overlay. v3 shuffles slot assignment independently, draws every choice from its own hash, and enforces label-aware visibility for gold AND negatives. Do not use v1/v2.
| field | type | meaning |
|---|---|---|
query |
str | the MCQ instruction (4 surface variants; names the product and the defect type; asks for the letter only) |
image |
Image | the 2×2 composite (~1560px, JPEG) |
annot |
str | gold letter A–D (exact match) |
reasoning |
null | none — items are deterministic; no teacher was used |
cate / task |
str | B / T-B2 (unified schema) |
metadata |
str (JSON) | source, category, image_sha256, d15_record_id, defect_type, bbox_xywh (native px of the source image), area_pct, gold_letter, panel_tags (letter → construction tag), equalized_base |
Provenance
Built from AI4Manufacturing/D15 (DefectSpectrum — EnVision-Research, ECCV 2024, arXiv:2310.17316;
fine-grained multi-class semantic re-annotation of MVTec-AD / VISION / DAGM / Cotton-Fabric), after D15's
2026-07-08 correction (26 upstream-misfiled records removed). Generator:
annotate/D15/build_d15_mcq.py in AI4Manufacturing/forge_model
— fully deterministic (seeded by md5(record_id|type)), zero API cost. Upstream license: MIT (respect the
underlying datasets' terms; this card is license: other).
Overlap / de-duplication (§8)
The base photos are the SAME images as D15 and therefore inherit all of D15's overlaps (sha-verified):
DS-MVTec ⊂ D20 test (and appears in
D05); DS-DAGM ⊂
181 (120 of them in 181's test);
DS-VISION ⊂ D23 (incl. its val split).
Do not evaluate on those repos' held-out splits if you train on D15-mcq. Reconstruct exact overlap sets via
metadata.image_sha256.
Training notes
- Composite images are self-contained — no extra rendering needed at train time.
annotis a single letter → put loss on completions; exact-match reward for RLVR.- Companion sets:
D15-annotated(CoT defect typing on the raw images, L1–L3 of the same ladder), and the sourceD15for pixel-mask GT.
- Downloads last month
- 19