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 filled reasoning column and this repo is not itself a training view. Derived repos each state their own regime on their own card.

216

Weld-surface defect detection on a MAG robotic welding line (4 author-named classes; native-pixel COCO bbox; two resolution tiers). 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

3,022 records (test=725 · train=2297).

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: one line per defect, `<class>,[x, y, width, height]` — native-pixel COCO xywh at the image's own resolution, top-left origin, NOT xyxy and NOT normalized. Four classes, the authors' own names: `pore`, `deposit`, `discontinuity`, `stain`. A bead with no annotated defect answers `none` (77 of them, all in the low tier). Each box also carries `short_side_px` and `legible_at_1x` in `metadata.objects[*]` — read the legibility section before training
`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

Licence — three layers, stated separately so each can be checked

1. The authors' own grant. Verbatim from the "Software license agreement" section of the project README (https://github.com/SylvioBlock/LoHi-Weld, checked 2026-09-09):

"Our dataset/code can be used for research, non-comercial or comercial purposes for free with proper citation." (the misspelling is the original's)

That permits commercial use, not merely research — broader than most sources in this corpus.

2. There is no LICENSE file. The GitHub API reports license: None and the repository root lists no licence file; LICENSE, LICENSE.md and COPYING all return 404. ⚠ The 404s are corroboration only — a deliberately bogus control path returns 404 identically — so the evidence is the API field and the root listing. Any GPL-3.0 association is inherited from the repository being a fork of WongKinYiu/yolov7, whose description it still carries; no GPL text exists in the repository.

3. Redistribution. The README grants use and says nothing about redistributing a converted copy. Permission for this converted copy was obtained from the authors and recorded on 2026-09-09 in the project's grant records. That permission is ours and is not transitive — it does not travel to anyone who redistributes this repository further.

Cite: S.B. Block, R.D. da Silva, A.E. Lazzaretti, R. Minetto, "LoHi-WELD: a novel industrial dataset for weld defect detection and classification, a deep learning study, and future perspectives", IEEE Access, 2024, DOI 10.1109/ACCESS.2024.3407019.

Provenance — three rules that change what you get if you ignore them

1. Labels come from the git repository at pinned commit 14e7b7a628af00c16bde27c58a032adc5c7c7802, not from the Drive zip. The repository carries all 3,022 .yolo label files and both k-fold trees, so the annotation is pinnable to a sha, which a Drive zip is not. The zip's copies of those labels were checked and are byte-identical — 0 of 3,022 differ.

2. Pixels come from the Drive zip, the only channel that offers them.

3. ⚠ Boxes are denormalized by the JPEG's ACTUAL size, never by the .json sidecar's declared size. The sidecars carry a real bug: on the low tier, imagePath drops the _0/_1 crop index, and 1,000 of the 2,000 low .json files declare an imageWidth/imageHeight that is not their own image's. Measured over every box in the dataset:

denormalized by	boxes reproduced
the JPEG's actual size	22,412 / 22,412 = 100%
the JSON's declared size	17,186 / 22,412 (low tier only 59.6%)

The data is sound; the sidecar metadata is not. This converter never reads the .json.

Box trimming: 260 boxes (1.2%) extend past the image edge in the source and are trimmed to it. Ten of those cross the 16 px line, which is why the shipped sub-floor rate for the low tier is 93.2% while the read-only audit — which measured the untrimmed box — reported 93.1%.

Two resolution tiers in one canon

The source ships two acquisition sets and both are here, boxes as measured, distinguished by metadata.resolution_tier:

tier	images	boxes	image size (median)	zero-box images
`high`	1,022	9,475	779 × 192	0
`low`	2,000	12,937	174 × 40	77
total	3,022	22,412		77

⚠ The images are crops of the weld bead, not camera frames. The paper's "640×480 / 2048×1080" are the camera resolutions. The crops were verified to be native-resolution cutouts: one source frame was downloaded and its crop located inside it by exhaustive search at scale 1, matching to MAE 0.09 grey levels. So the full-resolution originals the project also distributes add background context, not detail — they cannot recover a single pixel of defect resolution.

Class counts differ sharply between tiers — pore is 26.5% of low boxes but 5.5% of high:

class	high	low
stain	4,181	4,126
discontinuity	2,977	4,243
deposit	1,794	1,141
pore	523	3,427

⚠⚠ Legibility, and what you must do about it before training

Short side of every defect box, in native pixels, against the corpus's 16 px floor. Every box carries its own short_side_px and legible_at_1x in metadata.objects[*], so this is auditable per record rather than only in aggregate.

tier	boxes	p10	median	p90	under 16 px at 1×
`high`	9,475	11	24	55	28.2%
`low`	12,937	4	9	15	93.2%

Per class, median short side and share under the floor at 1×:

class	high	low
discontinuity	37 px — 0.6%	12 px — 86.3%
stain	19 px — 35.9%	9 px — 96.6%
deposit	17 px — 45.7%	9 px — 86.2%
pore	14 px — 63.5%	5 px — 100.0%

Every one of the low tier's 3,427 pore boxes is under the floor at 1×.

The adapt-layer instruction — read this before building a training view

The corpus's legibility floor is defined on the rendered short side, not on source pixels. The low crops are only ~174 × 40, so a processor left at its defaults feeds them at roughly 1×, and at 1× the low tier is not learnable as detection.

Do not put the low tier into detection training at 1×.
A default adapt view should render the low tier at ≥3× and recompute legibility at the render scale it actually uses. At 3× the low tier's residual sub-floor share falls from 93.2% to 22.3%; the median box goes from 9 px to 27 px. (The high tier already clears the floor at its median: it needs 0.7×, and at 3× only 0.2% of its boxes remain sub-floor.)
Upscaling adds no information. It is worth doing anyway because the measured damage comes from illegibility, not blur — a same-kernel blur control scores .994 while genuine 4 px illegibility drops good-recall from .770 to .482. Rendering larger lets the vision tower's patches resolve what is already there.

Canon measures and discloses; the render scale is a layer-3 decision, which is why nothing is filtered here.

⚠ `OK` / `NG` in the low filenames is NOT the label

It is the production line's own pass/fail verdict for the part, and it does not predict defects:

filename tag	has ≥1 defect box	has none
`NG`	1,103	37
`OK`	820	40

820 of the 860 OK images carry at least one annotated defect. It rides in metadata.line_verdict for anyone studying line-verdict agreement, and is never the answer. The two crops of a capture always share the tag (0 captures disagree).

Split — session-wise for `high`, capture-wise for `low`, never image-wise

The source's own protocol splits per image (create_kfold.py: train_test_split over the image array, then KFold(5)), and images share physical units, so it leaks:

low: 2,000 images are 2 crops each of 1,000 captures; the authors' folds put sibling crops on opposite sides — 2,290 capture collisions across the five folds.
high: 1,022 images come from 5 sessions (381 / 316 / 271 / 49 / 5); all five appear on both sides of every fold.

⚠ These are correlation leaks, not duplication leaks — a distinction worth keeping. All 3,022 images have distinct sha256, and inspection confirms the _0/_1 crops are two different weld beads from one capture, while consecutive high images are different short welds from one session, not repeated frames. What is shared is workpiece, parameters, lighting and time.

This repository therefore carves its own:

tier	rule	test	share
`high`	hold out whole sessions 27, 28, 29	325 images	31.8%
`low`	hold out 200 of the 1,000 captures, both crops together	400 images	20.0%
repo	the two tiers' test sets merged	725 images	24.0%

All four classes appear in both tiers' test sets. metadata.resolution_tier separates them again for per-tier evaluation.

The authors' k-folds are preserved in metadata.author_kfold (train_folds, val_folds, in_author_test) for anyone reproducing their protocol, each carrying a warning field recording that they are image-wise and leaky.

Lazy-baseline floors (test) — report against these, not against zero

blind predictor	score
always "there is a defect"	97.7% of test images are correct
always `none`	2.3%
always the train-majority class (`stain`)	37.7% of test boxes
always the train median count (7 boxes)	mean \|count error\| 3.69 boxes
pixel-blind probe: predict the tier's train mean count from `resolution_tier` alone	mean \|count error\| 3.44 boxes

The tier is stated in the record, so the probe is the floor a model beats only by looking at the bead. Presence is nearly vacuous here (97.7% of test images contain a defect), so localization and class are where the signal is.

Provenance

Underlying dataset: LoHi-WELD. Upstream license: Free for research, non-commercial and commercial use with citation, per the authors' README; no LICENSE file exists (GitHub API reports license: None). Redistribution of this converted copy was permitted by the authors, recorded 2026-09-09 — not transitive. See the licence note below (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 216/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Converter: forge_model@c564c18, merged in PR #89 as 35d4de8. That is the last commit to touch this dataset's converter, which is what produced the data; this card's own text lives in publish/push_to_hf.py and moves independently.

Overlap / de-duplication (§8)

No overlap with any other dataset in this corpus: 0 shared sha256 against 102 (RIAWELC, 24,407 images scanned) and 214/215 (GDXray, 46,608 scanned), and this is visible-light surface photography rather than radiography. ⚠ Two physical grouping keys — metadata.capture_id (low tier) and metadata.session (high tier) — and any re-split must respect them; the authors' own k-folds are image-wise and leak (see the split note below). Each record carries metadata.image_sha256 so overlapping images can be kept entirely on one side of a train/eval split.

Downloads last month: 12

Total file size:

119 MB

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