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

Roles

Roles: canon repo — annot is the source label, kept machine-parseable as the gold for verification and reward parsing; there is no reasoning column and this repo is not itself a training view. Derived repos (-annotated, -grounding, -region, -mcq) each state their own regime on their own card. Geometry for every record lives in metadata.geometry (below).

182

Multi-modal synthetic candy AD (8 categories; multi-label; all 8 modalities per record). Category B, task T-B2, in the unified Smart-Manufacturing SFT schema.

The repository name is an internal task code. See Provenance below for the underlying dataset.

Records

9,200 records (test=400 · train=8000 · validation=800). Pixel masks are embedded as a mask image column.

Unified SFT schema

field	type	meaning
`query`	str	the question / instruction (model input)
`image`	Image	the input image (bytes embedded); for multi-image rows, a preview of the first view
`images`	list[Image]	(multi-image rows) all input views / modalities for the row, bytes embedded
`annot`	str	the answer — for this dataset: plain-text multi-label `{label, [defect_types]}` — `{good, null}` or `{anomalous, [<type>, ...]}` over bumps/colors/dents/normals (a single object can carry several types, so the types are a bracketed list), from each sample's metadata flags. One record = one query over all 8 modality images of the same candy: the `images` sequence column holds them in a fixed order — 6 RGB lightings (image_0..5), depth, normals (`metadata.modalities` gives the per-position kind); the `image` scalar is image_0. The combined anomaly `mask` is deferred GT — see Task, modalities, mask & split below
`reasoning`	null	no native CoT in these datasets
`cate`	"B"	SFT category
`task`	"T-xx"	unified task id
`metadata`	str (JSON)	split, provenance, `image_path`, `image_sha256` (dedup key)
`mask`	Image \| null	(T-B1/T-B2 only) the pixel ground-truth mask, bytes embedded
`masks`	list[Image]	(multi-image T-B1 / D21) per-view masks aligned with `images` (None where a view has no defect), or multi-region masks

Task, modalities, mask & split

What this is. Eyecandies (Bonfiglioli et al., "The Eyecandies Dataset for Unsupervised Multimodal Anomaly Detection and Localization", ACCV 2022) — a synthetic (Blender-rendered) candy anomaly-detection dataset. 8 candy categories; each object is rendered under 6 lighting conditions (image_0..5) plus a depth map and a normals map, and defects carry pixel masks. Four defect types: bumps, colors, dents, normals (multi-label).

Task & answer. Multi-label defect classification + localization. The dataset ships no natural-language query (only per-sample anomaly flags in metadata.yaml), so query is our own template: it names the candy category and asks whether it is good or anomalous and, if anomalous, to list every defect type present. annot is {label, [defect_types]} — a single image can carry several defect types, so the types are a bracketed list: {good, null} / {anomalous, [bumps]} / {anomalous, [bumps, colors, dents, normals]}. Labels come from each sample's metadata.yaml flags. The query does not ask for a mask.

All modalities are embedded — one query over 8 images. Each record bundles all 8 modality images of the same candy so the model judges from the full multi-modal view. The images sequence column holds them in a fixed order: 6 RGB lightings (image_0..5), then depth, then normals — metadata.modalities lists the kind at each position and metadata.n_images the count (8). The image scalar column is image_0 (the primary RGB, for dataset viewers; its bytes are deduplicated against images[0] in the parquet). Depth is a 16-bit map and normals is a surface-normal map; both are stored as images. (Earlier revisions shipped the extra modalities as loose assets/ files — they are now embedded in images instead.)

Mask (deferred GT). The combined anomaly mask is the deferred localization ground truth in the mask column (anomalous images only; good = null); metadata.defect_area_fraction gives its area. A text model cannot emit a pixel mask, so segmentation is deferred.

Split (why anomalies are only in test). Eyecandies is an unsupervised anomaly-detection benchmark, so by design the training data is normal-only: train and validation contain 100% good samples and anomalies appear only in test. This is the source's own composition (verified from every sample's metadata.yaml flags), not a conversion artefact. The source's test_private split ships no masks/labels (GT withheld) and is dropped. Counts: train 8,000 good, validation 800 good, test 400 (210 good / 190 anomalous) = 9,200; the private test (3,200) is excluded.

Query text — pooled paraphrases (v2)

Every record's query is drawn from common/vision_query_pools.json[182/label_types_multiview], a pool of 37 gate-verified paraphrases of the shipped wording, assigned by a stable hash of the source image path and recorded as metadata.query_template (37 templates in use, top share 3.0%).

Template 0 is v1's wording byte for byte (266 records keep it); the pass asserted that on every record before rewriting anything.

Template ↔ gold independence on this build: 9,200 records, 37 templates, worst template p = 0.0126, alpha 2.7e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 9,200 images share one frame size.

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).

Provenance

Underlying dataset: Eyecandies. Upstream license: other (research use; Eyecandies, Bonfiglioli et al. ACCV 2022) (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 182/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Overlap / de-duplication (§8)

Synthetic (Blender-rendered) — no image overlap with the real-image AD sets.

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	82,800
distinct by decoded pixels	73,600
images carrying more than one record	0
images on both sides of the split	0

images checked counts every entry of images plus the image preview column, which is a copy of the first view — 9,200 records, 82,800 decoded images; the distinct count is therefore lower by one per record by construction, and only the more than one record row measures duplication.

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.

Coordinates are native pixels of the image in that record (coords_frame: "record_image"). scale is 1.0 throughout — this repo publishes at source resolution, nothing was downscaled at publish time.

"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:0`
`connectivity`	`4`
`merge`	`mask_dilate:1pct`
`min_area_px`	`15`
`max_instances`	`None`
`artifact`	`fine`
`fill_floor`	`None`
`legibility_floor_px`	`None`
`min_side_floor_px`	`None`
`spec_sha`	`5cdc1f4dbc364476`

Provenance and verification


records	9,200
carrying a geometry block	9,200 / 9,200
instances per record	`0`: 9,010, `1`: 149, `2`: 7, `3`: 13, `4`: 19, `5+`: 2
total instances	289
image dimensions	512×512 (9,200)
`scale` values present	[1.0]

Computed from this repo's own 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)	512×512 → 504×504
shipped boxes	289
legible at that render (>=16px there)	245 (84.8%)

⚠ 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 512×512 is rendered 504×504 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.

Downloads last month: 79

Total file size:

18.5 GB

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