This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.

186-mcq

Mask-grounded multiple choice (Set-of-Mark style) for magnetic-tile defect localization — 244 items, derived deterministically from the pixel saliency masks of AI4Manufacturing/186. Exact-match gradable → SFT and RLVR-ready.

Task

One item per eligible defective record. The image is a 2×2 grid of views A–D of the same tile photo; each view overlays one red candidate region mask (SoM style: translucent fill + outline + corner letter). Exactly one view overlays the true defect mask — in both location and extent (the query says so). annot is the correct letter. The three negatives per item are hard by construction, drawn from: shift (translated by ~0.7–1.6× bbox extents), fliplr/flipud/rot180 (mirrored), dilate (over-grown ≥2.5×, bounded), erode (shrunk <60%). This dataset has ONE defect type per image, so there is no othertype negative. Every negative is guaranteed wrong (IoU vs truth < 0.35 except dilate, wrong by extent) and panels are mutually distinct (pairwise IoU < 0.7). Negative kinds are assigned to slots by an independent salted hash (arrangement encodes nothing).

Gold letters: A 53 / B 61 / C 58 / D 72 (chi² = 3.18, at chance). Query pool: 16 variants.

Exclusions (counted, confidence-over-coverage): 148 of 392 defective records were skipped — gold overlay under ~30 visible px after panel downscale (37, mostly the smallest Blowholes), fewer than 3 sound visible negatives constructible (77), mask covering >35% of the frame (30, large Uneven/Fray — no sane negatives), source empty-mask Uneven rows (4). Per-type coverage of the 244 shipped items: Blowhole 100, Break 55, Crack 56, Uneven 20, Fray 13. The skipped defects remain fully covered by the companion 186-grounding / 186-region sets.

field	type	meaning
`query`	str	16 variants; names the defect type to locate; "answer with the letter only"
`image`	Image	2×2 composite, panels A–D
`annot`	str	`A` / `B` / `C` / `D`
`reasoning`	null	none — deterministic derivation
`cate` / `task`	str	`B` / `T-B2`
`metadata`	str (JSON)	source, category, `image_sha256`, `image_path`, `r186_record_id`, `defect_type`, `gold_letter`, `panel_tags`, `area_pct`

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 deterministically (no LLM/teacher; reasoning is null) from AI4Manufacturing/186 (revision 2117f8e) — Magnetic-Tile-Defect, Huang et al., "Surface defect saliency of magnetic tile", The Visual Computer 2020: 1,344 grayscale magnetic-tile images, 5 defect classes (Blowhole, Break, Crack, Fray, Uneven) + good, each defective image with a paired pixel saliency mask (binarized here at gray>40, which matches the source defect_area_fraction). Generator: annotate/186/build_186_derived.py in forge_model; machine gates: annotate/186/verify_186.py (all green at build time).

Source-data exclusion (counted): 4 MT_Uneven rows ship ALL-ZERO masks in the source dataset (defect_area_fraction = 0.0) — an anomalous label with no localizable GT. They are excluded from every derived set.

Query diversity. The query field is drawn from a fixed pool of surface variants for this task (paraphrases preserving the task and answer format), selected by an independent per-record hash. A machine gate checks that no template correlates with the gold (worst z-scores reported above).

The repository name is an internal task code (the source dataset's code is 186).

Geometry (`metadata.geometry`)

Every record carries a geometry block inside the existing metadata JSON string, so that its gold can be re-derived at any render size. No schema column changed; existing loaders are unaffected.

⚠ This repo's images are COMPOSED (a multi-panel montage), so the image in a record is not the image the coordinates belong to. The block says so explicitly:

image_wh — dims of the montage actually stored in the record
source_wh — dims of the source image the boxes are measured in
coords_frame: "source_image" — bbox_xywh are SOURCE pixels, not montage pixels
scale: null — a composed image has no single scale to the source

244 of 244 records are composed. The boxes are the regions the question was built from; they do not locate anything in the montage, and rescaling them by the montage size would be wrong.

"geometry": {
  "image_wh":  [W, H],        // dims of the image in THIS record
  "source_wh": [W, H],        // dims of the original source image
  "scale": 1.0,               // image_wh / source_wh; < 1.0 would disclose a publish-time downscale
  "n_instances": 2,
  "instances": [
    { "instance_id": 1, "bbox_xywh": [x, y, w, h], "min_side_px": 65, "class": null }
  ],
  "n_dropped_subminimum": 0,  // components removed by the filters below
  "union_box_fallback": false,// true => boxes are per-class unions, NOT real instances
  "conventions": { ... }      // see table
}

instances is present even when empty. [] means the record genuinely has no defects; an absent block would mean geometry could not be recovered. Those are different states and are never conflated.

Conventions used to derive it

There is no universal definition of "one defect instance" — it depends on the mask the source shipped. This repo's is stated, not implied:

field	value
`algorithm`	`dilate_cc`
`binarisation`	`gt:40`
`connectivity`	`4`
`merge`	`mask_dilate:1pct`
`min_area_px`	`15`
`max_instances`	`8`
`artifact`	`coarse`
`fill_floor`	`None`
`legibility_floor_px`	`None`
`min_side_floor_px`	`None`
`spec_sha`	`22cd9e70b8008b05`

Provenance and verification


records	244
carrying a geometry block	244 / 244
instances per record	`1`: 209, `2`: 28, `3`: 1, `4`: 3, `5+`: 3
total instances	295
image dimensions	1216×544 (3), 1222×512 (3), 1212×508 (3)
`scale` values present	[None]

Derived from the AI4Manufacturing/186 masks and verified against this repo's own published answers before it was written — a recomputation that disagreed with the shipped gold would have aborted the update rather than overwritten it.

⚠ The 16px floor applies at the RENDER, not at native

min_side_px is in native pixels. The model does not see native: Qwen2-VL caps by megapixels AND snaps each dimension to a multiple of 28. So min_side_px >= 16 is the floor tested in the wrong frame. Measured on this repo:


native → rendered (qwen2_vl @ 2.36MP)	234×590 → 224×588, 250×598 → 252×588, 250×600 → 252×588
shipped boxes	295
legible at that render (>=16px there)	134 (45.4%)

⚠ An earlier version of this section reported the inverse — boxes clearing 16px at native and failing at the render — and that number was misleading. It is frame-relative: publishing at a larger native size lets more boxes clear 16 in the published frame, so more can "fail", which penalises exactly the choice that helps. Measured on 179: publishing native (3024) means a box needs >=32px native to be legible at the render and 86.7% qualify; the previous 1024 publish needed >=47px native and only 69.5% qualified. The native republish improved rendered legibility by 17 points while the old metric scored it as 12.5% "broken". The figure above is the comparable one.

Nothing in the data is frame-dependent — geometry is native and complete. Use forge_model/common/adapt_engine.py, which applies the floor at whatever size the consumer renders.

Using it

Coordinates only stay correct if they are rescaled with the image. A patch-based VLM does not render at native size: Qwen2-VL's processor snaps both dimensions to a multiple of 28, so this repo's 234×590 is rendered 224×588 and native-pixel boxes are then wrong by a few pixels. forge_model/common/adapt_engine.py regenerates coordinates for a target render size, re-derives counts, and drops records whose gold no longer holds there.

Query text — pooled paraphrases (v2)

Every record's query is drawn from common/vision_query_pools.json[186/mcq], a pool of 33 gate-verified paraphrases of the shipped wording, assigned by the rung builder's own hash of the derived record id and recorded as metadata.query_template (33 templates in use, top share 4.9%).

v1 drew from the first 16 of these templates with the same hash rule (7 records keep their v1 text). Query pool only: the rung's pool was topped up to the 30-variant floor (originals first, same hash rule as the builder, so a rebuild from source gives the same text); annot and reasoning are untouched (reasoning is null on every record of this rung).

Template ↔ gold independence on this build: 244 records, 33 templates, worst template p = 0.0268, alpha 3.0e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): balanced accuracy 0.189 vs 0.250 chance (plain 0.197 vs 0.328 majority), 228 distinct frame sizes — no signal (5-fold cross-validation: the repository ships one split).

Answers, images, masks, split and every other field are byte-identical to v1: this revision was issued from the published parquet itself (tools/requery_published.py), not rebuilt from source, and the pixel-identity guard ran on the embedded images (§8 below).

Image identity — measured at this republish

Two identities, and they answer different questions. metadata.image_sha256 hashes the file bytes: it finds byte-identical copies and is blind to a re-encode. metadata.pixel_sha256 hashes the decoded image (mode | size | pixels): it finds the same photograph saved twice. Only the second one settles whether an image is duplicated.

Measured at build time, not asserted afterwards — a violation aborts the build and names the offending records:


images checked	244
distinct by decoded pixels	244
images carrying more than one record	0
images on both sides of the split	0

Downloads last month: 60

Total file size:

54.6 MB

You need to agree to share your contact information to access this dataset