Datasets:

AI4Manufacturing
/

D15-region

tags:
  - smart-manufacturing
  - sft
  - industrial
  - vision
license: other
pretty_name: D15-region
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
extra_gated_prompt: >-
  This dataset is released for **research use**. Access is reviewed and granted
  **manually** by the maintainers. Please state your name, affiliation, and
  intended use.

D15-region

Region-conditioned defect typing (L5 of the D15 task ladder) — 6,040 items, derived deterministically (no LLM/teacher) from the semantic masks of AI4Manufacturing/D15 (DefectSpectrum). Exact-match gradable (closed type list + no defect) → SFT and RLVR-ready.

The repository name is an internal task code. See Provenance below.

Query diversity (2026-07-11). The query field is drawn from a pool of 40 per mode (80) 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 in verify_*.py checks that no template correlates with the gold (binomial z < 4.5).

Task

"An operator points at a region — what defect, if any, is there?" One item per defect instance (per-class connected components after proximity grouping, so fragmented defects like dashed scratches form one instance; 107 over-fragmented records fall back to per-class union boxes — metadata.union_fallback marks union-box items) plus 1,490 clean-region negatives (~25%) teaching rejection. The region is conveyed in one of two modes (~50/50, metadata.region_mode):

overlay — a red rectangular ring drawn on the image around the region (ring thickness scales with image size; regions are padded to ≥6% of the min dimension so they stay visible; positives and clean regions use the identical rectangular ring language, so ring shape leaks nothing).
bbox_text — the raw image plus the region as a native-pixel box [x, y, w, h] (origin top-left) in the query text — pairs with the coordinate vocabulary of D15-annotated's reasoning_grounded.

Golds are guaranteed unambiguous: instances whose query box contains another defect class's pixels are skipped (422), and clean boxes contain zero defect pixels by construction. Gold = the type name exactly as listed in the query's closed class list, or no defect.

Records

6,040 items (single train split): 4,550 positives + 1,490 clean. Top golds: no defect 1,490 · scratch 1,220 · crush 643 · color 228 · bump 213 · crack 202 · … Counts auto-generated by annotate/D15/verify_d15_family.py --card-numbers (2026-07-17).

Excision (2026-07-17): one contradiction record removed (image sha256 4be393b8..., DS-VISION/Capacitor/512.jpg — source labels it anomalous with an empty mask), which contributed one overlay/no defect item here. The base D15 repo retains it as a disclosed source contradiction.

Per-category prior warning: per-category modal-answer shares reach 70.8% (Screw→crush) — see the Training-mixture notes below; the disclosed soft geometry priors remain as documented.

Revision notes (2026-07-09): two adversarial review rounds. v1→v2: template-parity tell (two templates only occurred on defect items) + two defect types lost entirely to the sub-15px denoise (union-box rescue added, metadata.union_fallback). v2→v3: clean boxes were size-clamped to [6%, 30%] while defect boxes weren't — printed coordinates separated defect from clean with 100% precision on 77% of bbox-text items. v3 samples clean-box sizes from the record's own defect boxes (no clamp; overlay minimum applies to both kinds), jitters good-record sizes, and guards the drawn ring's full visual extent for purity. v3→v4 (round-3 review): clean sizes previously sampled per-CLASS union boxes while positives used per-INSTANCE boxes — a within-category size tell; v4 samples clean size AND position from the exact per-instance population positives emit. v4→v5 (round-4 review): v4's clean sampler had an 8-px size floor and off-by-one edge clamps — coordinates only positives could print (336 bbox-text items decodable at precision 1.0). v5 removes every clean-only constraint (no floor, flush-to-edge allowed), draws clean sizes/centers from the exact emitted-positive box population (instances + union boxes), and adds a uniform placement fallback so the size/position tail stays reachable. Measured on v5: bbox_text geometry GBM AUC 0.51 (chance); per-category size-only probes at/below majority everywhere; out-of-range rule scan within bounds below. Known residuals (inherent, disclosed, regression-gated in verify_leaks.py — round-5 adjudication; no precision-1.0 rule exists in v5, verified): (1) an overlay position prior (global GBM AUC ~0.64): clean regions physically cannot occupy defect pixels — e.g. zipper defects lie on the zipper line; (2) extreme-tail marginals in bbox-text (54 area + 26 x items, ~2.5%): the SMALL-area tail is ~13× defect-heavy because every defect instance emits a positive while each record draws at most one clean (pool weighting); the far-edge/giant tail is purity-constrained (several sources, e.g. VISION, contain no defect-free stock); (3) a large-box position×size interaction (boxes ≥10% of the image, ~7% of items: position-only AUC ~0.74–0.80 within that stratum): large clean boxes place uniformly while large defects sit at object-structured positions — the irreducible trade for tail coverage (the alternative, no large cleans at all, is a strictly worse precision-1.0 rule). All are soft priors from purity-constrained negatives — none bypasses looking at the image, and the region CONTENT remains the intended signal. Do not use v1–v4.

field	type	meaning
`query`	str	4 surface variants per mode; names the product; closed class list; instructs "type only, or 'no defect'"
`image`	Image	product photo, with the red ring burned in for `overlay` mode
`annot`	str	gold type string, or `no defect` (exact match)
`reasoning`	null	none — deterministic derivation
`cate` / `task`	str	`B` / `T-B2`
`metadata`	str (JSON)	source, category, `image_sha256`, `d15_record_id`, `region_mode`, `bbox_xywh` (native px; padded box for overlay), `instance_index` (−1 = clean), `gold`

Roles

Roles: this is an answer-only tier — there is no reasoning column; annot is both the machine-parseable gold AND the direct-answer SFT target ('SFT-ready' here means direct imitation of annot in the query-specified format); it is also the exact-match/IoU reward key for RLVR.

Provenance

Built from AI4Manufacturing/D15 (DefectSpectrum, ECCV 2024, arXiv:2310.17316) after its 2026-07-08 correction. Generator: annotate/D15/build_d15_l5_grounding.py in AI4Manufacturing/forge_model — deterministic (seeded by record id), zero API cost. Upstream license: MIT (respect upstream terms; card is license: other).

Training-mixture notes (2026-07-17 certification)

One photo appears in up to 5 family artifacts (family mean ≈6.3 items/photo). Carve any train/eval split PHOTO-WISE on metadata.image_sha256 across ALL D15-family repos simultaneously.
Global label prior: 85.6% of photos are defective — a blind guesser scores ≈90% on good-vs-defective labels. Counterweight with good-heavy sources in your mixture; this is a structural property of DefectSpectrum, not a leak (text-only blind probe on shipped queries: MCQ at exact chance 25%; region at prior level).
641 VISION photos in this family are byte-shared with AI4Manufacturing/D23 under a MATERIALLY DIFFERENT label policy (DS masks confirm only 88.5% of VISION boxes; worst subset coverage 0.03). Take each shared image from exactly ONE side — machine-readable keys: overlap_with_D23.json on the D15 repo. Full sibling-overlap table: D15 base card §8.

Overlap / de-duplication (§8)

Same base photos as D15 → inherits all of D15's sha-verified overlaps: DS-MVTec ⊂ D20 test (also in D05); DS-DAGM ⊂ 181 (120 in its test); DS-VISION ⊂ D23 (incl. val). Do not evaluate on those repos' held-out splits if you train on this set. Reconstruct overlaps via metadata.image_sha256.

Training notes

Loss on completions; answers are short by instruction (the query demands "type only"), so no verbosity prior is learned. Exact-match reward for RLVR.
Companions: D15-annotated (CoT, L1–L3), D15-mcq (mask MCQ, L4), D15-grounding (detection format, L6-lite).