You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use. Access is reviewed and granted manually by the maintainers. Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
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-tiles
Tile-level hot-crack detection in light-microscope images of a weld surface — the trainable view of 219 (512x512 native-resolution tiles; binary 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
6,694 records (test=1196 · train=5498). 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: one word, crack or no crack, for the 512x512 tile. The binary mask for the tile rides in the mask column as deferred localization GT. ⚠ This view must NOT be used for counting: 19.0% of crack components are cut by a tile boundary, so a split crack is counted twice — count on the whole frame in 219 |
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 |
What this is, and what it is derived from
The trainable view of the BAM hot-cracking data. Canon is
AI4Manufacturing/219 — 609 whole frames,
crack count, mask column — and canon's own card says plainly that whole-frame use is not viable:
at a long-side 1024 resize the median crack is 4.9 px and 96.1% fall under the corpus's 16 px
legibility floor.
This repository cuts those frames into 512 × 512 tiles at the microscope's native resolution, so a
median 17.2 px crack stays 17.2 px and becomes 3.4% of the tile's width — the same regime as
216's high tier (24 px in a 779 px image, 3.1%), i.e. an ordinary detection-scale problem.
⚠ This is optical microscopy, not production-line imagery
In the depositors' own words (Zenodo record 16919145): "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", taken "with an optical light microscope (Keyence)" at 50× and 100×. The weld surface of a test sample under a microscope — not a welding-line camera, and not sectioned metallography (a correction, 2026-09-12: the earlier wording "sectioned metallographic specimens / weld metallography" was an inference, not a statement in the record).
Licence
CC BY 4.0, from the Zenodo record's own metadata.license field. Cite the source deposit:
"Semantic Segmentation for Hot Cracking", Bundesanstalt für Materialforschung und -prüfung (BAM),
DOI 10.5281/zenodo.16919145.
Why 512, and why no overlap
Measured over all 609 masks before choosing, counting a tile as crack-bearing if it holds any foreground pixel:
| tile | stride | tiles | with crack | empty | coverage in a crack tile (median / p90) | components split |
|---|---|---|---|---|---|---|
| 512 | 512 | 18,065 | 3,447 (19.1%) | 80.9% | 0.42% / 2.39% | 19.0% |
| 512 | 256 | 65,439 | 13,478 (20.6%) | 79.4% | 0.43% / 2.44% | 33.0% |
| 768 | 768 | 7,263 | 2,537 (34.9%) | 65.1% | 0.27% / 1.45% | 13.5% |
| 1024 | 1024 | 3,582 | 1,918 (53.5%) | 46.5% | 0.20% / 0.98% | 9.0% |
The trade runs in opposite directions. A smaller tile makes the crack a larger share of what the model sees (0.42% at 512 against 0.20% at 1024 — the whole frame is 0.118%) but cuts more cracks at a boundary.
⚠ Overlap is the obvious knob and it is the wrong one. Halving the stride at 512 quadruples the tile count to 65,439, raises the split rate to 33%, and buys no coverage at all (0.43% against 0.42%). It is not used.
512 is chosen because it puts the crack back into an ordinary detection scale. At 1024 the crack falls to 1.7% of the tile — halfway back toward the problem tiling was meant to solve.
⚠ The table above counts any foreground pixel. This build requires a component of at least 50 px,
the same threshold canon 219's answer uses, so its shipped 512/512 numbers are 3,347 (18.5%) crack
tiles, 81.5% empty, coverage median 0.44% / p90 2.42%. The 100-tile difference is specks below the
threshold. Where the two disagree, the shipped numbers are the ones quoted below.
⚠⚠ This view must not be used for counting
1,227 of the 6,454 crack components (19.0%) are cut by a tile boundary, so a split crack appears
in two tiles, is counted twice, and is truncated in both. Counting belongs on the whole frame,
which is exactly what canon 219 answers. Here the answer is a tile-level verdict, with the mask
as deferred localization ground truth.
⚠ Two different denominators, not interchangeable:
| crack components cut by a tile boundary | 1,227 / 6,454 = 19.0% ← the no-counting argument rests on this |
| tiles showing a partial (cut) crack | 2,100 / 3,347 = 62.7% ← a different denominator |
One long crack crossing five tiles makes five tiles "partial" while being a single split component.
metadata.crack_truncated_by_tile_edge is the per-tile flag.
⚠ Empty tiles are sampled, and the sample is stratified
81.5% of tiles are empty. Shipping all of them would rebuild, at tile level, exactly the
95.7%-floor presence verdict that canon 219 rejected.
Every crack tile is kept; empties are sampled 1:1 → 3,347 + 3,347 = 6,694 records, a 50% majority floor.
⚠ The sample is stratified by (wire folder, magnification) (metadata.specimen × magnification),
and this is not a detail. Canon's crack-free frames are 23 of 26 from wire folder 1 and 25 of
26 at 50×; an unstratified empty sample would inherit that skew and teach the model that "wire 1 at
50× means no crack". Every stratum is filled exactly 1:1 with no shortfall:
wire folder (specimen) |
50× crack / empty | 100× crack / empty |
|---|---|---|
| 1 | 123 / 123 | 168 / 168 |
| 2 | 446 / 446 | 706 / 706 |
| 3 | 467 / 467 | 691 / 691 |
| 4 | 52 / 52 | 96 / 96 |
| 5 | 234 / 234 | 364 / 364 |
The shipped 50/50 balance is a sampling choice, not the true rate. The real rate — 18.5% of
tiles carry a crack — rides in metadata.true_crack_tile_rate so a consumer can re-weight.
Split
By wire, holding out wire 5 — the same folder canon 219 holds out, so the two repositories can be
used together without leaking across them. The record's structure is "Wire → Magnification (50, 100)
→ Data (train, mask = label) → Numbered images": the five top-level folders are filler wires, not
specimens, and one wire folder holds the surfaces of several test samples. Tiles from one wire are
not independent and a random tile split would leak far worse than the frame split does.
| records | crack | no crack | |
|---|---|---|---|
| train (wires 1–4) | 5,498 | 2,749 | 2,749 |
| test (wire 5) | 1,196 | 598 | 598 |
0 wires shared. Test is 17.9% of records. metadata.specimen keeps its values (= wire folder
1–5) and its name until the next data revision renames it.
Lazy-baseline floor
Majority answer → 50.0% on test (598 crack, 598 no crack). The balance is by construction, so
50% is the number to beat and accuracy is a fair metric here — unlike canon 219, where presence is
95.7% and had to be replaced by a count.
The image's width and height alone (the corpus's standing frame-size floor, common/lazy_floors.py):
vacuous by construction — every tile is 512 × 512, one frame size, so the probe has nothing to read
and scores 0.500 on both plain and balanced accuracy. This is the property canon 219 lacks: there,
511 distinct frame sizes reach MAE 4.39 on the count with no pixels read; here the tiling removes the
frame as a feature.
Each record carries specimen, magnification, tile_origin_xy, tile_px, stride_px,
crack_coverage_frac, crack_truncated_by_tile_edge and true_crack_tile_rate, so every choice
above is auditable per record rather than only in aggregate.
Version history
v1 — published 2026-09-09. 6,694 records (5,498 train / 1,196 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. Every record's query is drawn from a pool of 37
gate-verified paraphrases of the v1 wording, assigned by a stable hash of the tile's key and recorded
as metadata.query_template; index 0 is v1's wording, byte for byte (188 records keep it); the
magnification slot makes 74 distinct query strings. Answers, tiles, masks, split and every other field
are unchanged. The pool passed the deterministic gates (names the target, keeps the two-word answer
form and the exact answer strings, 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: 6,694 tiles, 6,694
distinct, 0 on both sides of the split) and the frame-size floor line above.
v3 — 2026-09-12. Query wording and card only. Corrected against the Zenodo record: the top-level
folders are filler wires, not specimens (split unit = wire); the tiles are cut from weld-surface
light microscopy, not sectioned metallography. In the query, "test coupon" / "coupon" became
"test sample" / "sample" on every template, template index per record unchanged (slots recovered
through the template; 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. Tiles, masks,
answers and split are unchanged.
2026-09-12: section ⚠ This is optical microscopy, not production-line imagery rewritten in this revision; the previous text is superseded (2 of 2 lines replaced).
2026-09-12: section Split rewritten in this revision; the previous text is superseded (6 of 7 lines replaced).
Query text — pooled paraphrases (v2)
Every record's query is drawn from common/vision_query_pools.json[219-tiles/verdict], 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 3.2%).
Wording corrected in place (coupon → sample): 6,694 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: 6,694 records, 37 templates, worst template p = 0.0636, alpha 2.7e-04, 0 flagged → PASS.
Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 6,694 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: BAM hot-cracking weld-surface microscopy (512px native tiles). 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)
Derived from 219 (the 609 whole frames) and holding out the same wire folder 5 (metadata.specimen), so the two can be used together without leaking across them. No overlap with anything else in this corpus: weld-surface LIGHT MICROSCOPY of hot-cracking test samples, not production-line surface inspection like 216/217 or radiography like 102/214. ⚠ The grouping key is metadata.specimen — the filler-wire folder (the field keeps its name until the next data revision) — and any re-split must be wire-wise; the shipped 50/50 class balance is a sampling choice, not the true 18.5% rate.
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 | 6,694 |
| distinct by decoded pixels | 6,694 |
| images carrying more than one record | 0 |
| images on both sides of the split | 0 |
- Downloads last month
- 58