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.

219

Hot-crack counting in light-microscope images of the weld surface of Varestraint-Transvarestraint test samples (50x/100x; binary pixel masks). ⚠ Microscopy, NOT production-line imagery. Category B, task T-B1, in the unified Smart-Manufacturing SFT schema.

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

Records

609 records (test=152 · train=457). 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 (integer)	the answer — for this dataset: a single integer: the number of separate hot-crack regions, defined as connected components of the source mask with at least 50 px. Touching cracks count as one region and sub-50 px marks are not counted — both stated in the query, because a count is undefined without its referent. Derived deterministically from the mask; no teacher was involved. The presence verdict was rejected as the answer (95.7% majority floor, negative class from one wire folder) — see the card
`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

⚠ This is optical microscopy, not production-line imagery

What the images are, in the depositors' own words (Zenodo record 16919145, metadata.description):

"The images in this dataset are light microscope images of the surfaces of modified Varestraint-Transvarestraint test samples. The training images show the surface of a tungsten inert gas weld with possible hot cracks." — "The training images were taken in 24-bit, 3-color channels with an optical light microscope (Keyence). […] The images were taken at magnifications of 50x and 100x."

So: the weld surface of a hot-cracking test sample under a light microscope at 50× / 100× — not a camera on a welding line (the object is a test sample and the instrument is a microscope), and not a sectioned metallographic specimen: the record describes surfaces, and says some samples were only "prepared for further evaluation" after the surface images were taken. The cracks are surface hot cracks of a TIG weld.

Correction, 2026-09-12: until this revision the paragraph above called the images "sectioned metallographic specimens … weld metallography". That was an inference from "microscope", not a statement in the record, and it was wrong; the record's own sentences replace it.

Licence

CC BY 4.0, read from the Zenodo record's own metadata.license field. Redistribution permitted with attribution; no permission request needed. Cite the deposit: "Semantic Segmentation for Hot Cracking", Bundesanstalt für Materialforschung und -prüfung (BAM), DOI 10.5281/zenodo.16919145. The archive (segmented.zip, 11,427,069,320 B) was verified against Zenodo's published md5, recomputed locally, before anything was built from it.

⚠⚠ Whole-frame use is not viable — read this before training on it

Cracks cover a median 0.118% of an 8.1 MP frame at a median native width of 17.2 px. An 8.1 MP frame cannot be fed at native resolution, and every realistic resize destroys the target:

	native	long side 2048	1536	1024	512
median crack width	17.2 px	9.8	7.3	4.9	2.4
under the 16 px floor	42.7%	72.4%	84.9%	96.1%	99.8%

42.7% are already sub-floor at native resolution, and even at a long side of 2048 — already a large VLM input — 72.4% are under it.

This is the same damage the corpus measured on 216's low tier, reached from the opposite direction: there the image was too small, here the image is too large and the target is a hairline. The remedy differs too. 216's is an upscale, a dataloader flag. This one's is tiling, which has to cut the mask identically and choose an empty-tile policy — a converter's job, not a flag.

The trainable view of this data is AI4Manufacturing/219-tiles: 512×512 native-resolution tiles, where a 17.2 px crack stays 17.2 px and is 3.4% of the tile's width — the same regime as 216's high tier. This repository is the canon it is derived from: the source's own unit, with the damage measured rather than hidden.

⚠ Why the answer is a crack COUNT and not a presence verdict

The obvious image-level question — "is there a crack?" — was rejected, and the reason is recorded here so nobody re-derives it as the natural choice:

only 26 of 609 frames (4.3%) are crack-free, a 95.7% majority floor;
and the crack-free frames are not a sample of sound weld — 23 of the 26 come from wire folder 1 and 25 of 26 are at 50×. The negative class is one wire's artefact.

annot is instead the number of separate hot-crack regions: connected components of the source's own mask with at least 50 px, derived deterministically — no teacher was involved. This follows the corpus's counting rung (193), and the definition is stated in the query (touching cracks count as one region; marks under 50 px are not counted) because a count is undefined without its referent.


distribution	0 – 38 cracks, median 9, mean 10.6
majority-count floor	always answer 4 → 6.1% exact (against 95.7% for the binary)
median by wire folder (`metadata.specimen`)	1: 4 · 2: 13 · 3: 13 · 4: 17.5 · 5: 6
median by magnification	50×: 7 · 100×: 12

Lazy-baseline floors (test) — mean absolute count error

blind predictor	MAE
always 0	7.01 cracks
always the train median (11)	5.50
the train median for that magnification (50×→8, 100×→14)	5.22
⚠ the image's width and height alone (linear probe, no pixels read; `common/lazy_floors.py`)	4.39
(spread of the truth itself on the held-out wire)	sd 4.82

⚠ Magnification alone moves the answer, so it is stated in the query as a known condition. ⚠ And the frame's dimensions alone beat every blind floor above: the 609 frames come in 511 distinct sizes (2048 × 1536 up to 9478 × 4709 — nearly every frame its own), a bigger frame holds more cracks, and a linear probe on nothing but (width, height) reaches MAE 4.39 on the held-out wire — better than the per-magnification median (5.22). That is the floor a model must beat to have read the pixels at all. A model at MAE ≈ 4 may be doing little more than reading the frame size; real work starts below it.

Burned-in text. Every frame carries a "500 µm" scale bar in its bottom-right corner, stamped into the pixels by the microscope software. It was probed as a shortcut in its own right — a linear model on that corner alone, on this split — and carries nothing: crack presence 0.961 against a 0.980 majority floor (balanced accuracy 0.456 against 0.500 chance, class-balanced fit), crack count MAE 5.65 against the 5.50 blind median (ridge, λ = 10 on the 576-value corner patch; the frame-size row above is common/lazy_floors.py, λ = 1 on two features). The scale bar restates the magnification the query already gives; the shortcut in this repository is the frame size, not the text. Published images are never cropped (tools/burned_in_probe.py).

Structure and split


image/mask pairs	609, complete — 0 unpaired, 0 size mismatches
wire folders (`metadata.specimen`, values 1–5)	5 (124 · 162 · 153 · 18 · 152 pairs)
magnifications	50× (322) and 100× (287)
mask values	binary `{0, 255}`
image size	median 3516 × 2298 = 8.1 MP; range 2048×1536 – 9478×4709

The deposit ships no split — train/ is the image folder's name, not a division. The five top-level folders are filler wires, not specimens. The record states the structure as "Wire → Magnification (50, 100) → Data (train, mask = label) → Numbered images (1, 2, 3, …)" and explains the gaps in the numbering by "1) I did not take images of all surfaces if they were not cracked, and 2) the 100x magnification images were taken at a later point, by which time some samples had already been prepared for further evaluation" — so one wire folder holds the surfaces of several test samples welded with that wire (the record gives the wires no further identity than the numbers 1–5). This split is ours, and its unit is the wire: every frame under one wire folder shares the filler wire, and the numbered surfaces recur at both magnifications inside a folder, so no surface can straddle the sides. Wire 5 is held out (152 pairs, 25.0%). It was chosen because its count median (6) sits between the extremes (4 … 17.5), so it is neither the easiest nor the hardest wire; wire 4 (18 pairs) is too small to measure on and too odd to represent anything. 0 wires are shared between the sides.

metadata.specimen keeps its values (= wire folder 1–5) and its name until the next data revision renames it; metadata.split_note still carries the earlier "specimen" wording for the same reason. Both are data, and this revision changes no data.

Every record carries crack_count, crack_coverage_frac, crack_width_px_native and legible_at_1x, so the legibility measurement above is auditable per record and not only in aggregate.

Version history

v1 — published 2026-09-09. 609 records (457 train / 152 test), one query template. Its card's converter line read PENDING (PR #None): the card was rendered before the converter's PR existed and never re-rendered. Fixed here.

v2 — this revision. Query text only, and one word of it on purpose. v1's query asked for "the separate hot-crack regions annotated in it" — the word named the gold rather than the picture (a model is shown the micrograph, not the annotation). v2's base wording says "visible in it"; the count's referent is unchanged and still fully stated in the query (touching cracks are one region, marks under 50 px are not counted, one integer). No record keeps v1's wording — this is the one repository in the wave whose pool index 0 is not v1's sentence, and this paragraph is where that is recorded. Every record's query is drawn from a pool of 37 gate-verified paraphrases of the v2 wording, assigned by a stable hash of the image path and recorded as metadata.query_template (17 records carry the base wording); the magnification slot makes 74 distinct query strings. Answers, images, masks, split and every other field are unchanged. The pool passed the deterministic gates (names the target, states the counting rule and the integer answer form, never uses the gold word annotated, introduces no class name the source does not state) and a template ↔ gold independence test on this build (37 templates, 0 flagged). Two more things land with it: the build-time pixel-identity check (§8 now carries measured numbers: 609 images, 609 distinct, 0 on both sides of the split) and the frame-size floor row above, re-measured from the build itself (MAE 4.385, the same number the source-side probe gave).

v3 — 2026-09-12. Query wording and card only. Two statements were corrected against the Zenodo record (metadata.description at https://zenodo.org/records/16919145): the five top-level folders are filler wires, not specimens (the split unit is the wire); the images are weld-surface light microscopy, not sectioned metallography. In the query, "test coupon" / "coupon" became "test sample" / "sample" — the record's own word — on every template of the pool, template index per record unchanged (metadata.query_template as in v2; slots recovered through the template and the render asserted equal to the same correction applied to the v2 text). metadata.specimen retains its values (= wire folder 1–5) until the next data revision renames it. Answers, images, masks, split and every other field are unchanged.

2026-09-12: section ⚠ This is optical microscopy, not production-line imagery rewritten in this revision; the previous text is superseded (4 of 4 lines replaced).

Query text — pooled paraphrases (v2)

Every record's query is drawn from common/vision_query_pools.json[219/count], 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 4.4%).

Wording corrected in place (coupon → sample): 609 records changed, every record keeps its template index; the slots were recovered through the template and the render asserted equal to the same correction applied to the previous text. See the Version history.

Template ↔ gold independence on this build: 609 records, 37 templates, worst template p = 0.0596, alpha 2.7e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): MAE 4.385 vs 5.500 blind median, 511 distinct frame sizes — beats the blind median by 20%, report against it.

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: BAM hot-cracking weld-surface microscopy. Upstream license: CC BY 4.0 (Zenodo record metadata.license; Bundesanstalt fuer Materialforschung und -pruefung) — redistribution permitted with attribution (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 219/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Converter: forge_model@cae2e84 (PR #98). 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, and a different cell from the rest of the weld batch: this is weld-surface LIGHT MICROSCOPY of hot-cracking test samples (Zenodo record 16919145: "light microscope images of the surfaces of modified Varestraint-Transvarestraint test samples"), not production-line surface inspection like 216/217 or radiography like 102/214. ⚠ Whole-frame use is not viable — at a long-side 1024 resize 96.1% of cracks fall under the legibility floor; the trainable view is the derived 219-tiles. The grouping key is metadata.specimen — the filler-WIRE folder of the deposit (5 of them; the field keeps its name until the next data revision) — and any re-split must be wire-wise.

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	609
distinct by decoded pixels	609
images carrying more than one record	0
images on both sides of the split	0

Downloads last month: 66

Total file size:

11.6 GB