Datasets:
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.
102
Radiographic weld-defect classification over X-ray weld-bead tiles (4 classes; no boxes, no masks). 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
24,407 records (test=2443 · train=15863 · validation=6101).
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: the plain-text class name, exactly what the query asks for and nothing else — one of cracks, lack of penetration, no defect, porosity. The paper's abbreviation (CR/PO/LP/ND) and the upstream folder (Difetto1/Difetto2/Difetto4/NoDifetto) are in metadata, not in the answer. Whole-image classification — no mask, no boxes |
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, split & the leakage you have to handle
What this is. RIAWELC (Totino, Spagnolo, Perri, "RIAWELC: A Novel Dataset of Radiographic Images for Automatic Weld Defects Classification", Int. J. of Electrical and Computer Engineering Research 3(1), 2023, DOI 10.53375/ijecer.2023.320) — 24,407 radiographic tiles of weld beads in four classes: cracks (CR), porosity (PO), lack of penetration (LP), no defect (ND).
One record is a WINDOW, not a part. Per the paper's §II, the sources are 2000×8640 8-bit JPEG X-ray films shot in a real production environment. The weld bead was sliced out of each film, an 80×80-pixel window was swept over it, and each window was contrast-enhanced (CLAHE) and rescaled. Thousands of tiles therefore come from the same handful of films — which is what makes the split analysis below the most important thing on this card.
Counts reproduce the paper's Table I exactly, per class and per split — asserted in 102/convert_102.py,
so a drifted upstream release fails the build instead of publishing a wrong card.
⚠ The published tiles are 227×227, not the 224×224 the paper and the GitHub README both state. Measured on
all 24,407 files (grayscale mode L, PNG). metadata.image_wh is measured per record, never copied from the paper.
Query & answer. query is our own template (the source ships no natural-language question — only class
folders). It says what the image is, enumerates the four classes, and asks for the class name; annot is that
class name as plain text. One template is used for the whole dataset — the canon convention here; query
diversification belongs to the training-view builder (common/adapt_engine.py), not to the canon.
⚠⚠ The official split is leaked — upstream, and completely.
| records | 24,407 |
| distinct images | 21,964 |
| duplicate groups | 2,443, every one of size 2 |
| split pattern of every duplicate group | one train row + one test row |
| duplicate groups whose two members share the filename | 2,443 / 2,443 |
| duplicate groups with conflicting labels | 0 |
| test rows that are a byte-identical copy of a train row | 2,443 / 2,443 = 100.0% |
| validation rows byte-duplicated anywhere | 0 / 6,101 |
validation rows pixel-identical to a train row (v2, metadata.pixel_sha256 over the decoded pixels — a re-encode the file hash cannot see; same source film, overlapping strips, labels agreeing) |
25 / 6,101 |
| distinct images by decoded pixels (v2) | 21,893 (also 44 duplicate groups inside train and 2 inside validation, labels agreeing) |
The whole test split is a copy of part of the train split. This is the authors' own construction, not a
conversion error: Table I arranges train / validation / test as 65% / 25% / 10% of the collected images, and
the 10% test share was drawn without removing those images from the 65%. Verified twice and independently — from
the release archive's own file listing (the testing filename set minus the training filename set is empty) and
from SHA-256 over the decoded PNG bytes. De-duplicating the test split leaves nothing: 0 rows. The paper's
headline 93.33% test accuracy is measured on images the model was trained on.
⚠ And the validation split is not clean either — it shares source radiographs with train.
| grouping | groups | groups spanning >1 split | validation rows whose group also fed train | test rows likewise |
|---|---|---|---|---|
source_strip (one swept strip of one film) |
479 | 456 (95.2%) | 6,093 / 6,101 = 99.9% | 2,443 / 2,443 = 100.0% |
source_film (one X-ray film) |
109 | 108 (99.1%) | 6,101 / 6,101 = 100.0% | 2,443 / 2,443 = 100.0% |
Validation tiles are crop-disjoint from train, but they are neighbouring windows of the same weld on the same film. 71 of the 109 films carry more than one class, so a film-wise carve does not degenerate — it is simply not the carve the release ships.
Lazy-baseline floors and pixel-blind probes (report accuracy against these, never against 25%).
| split | n | majority class | majority floor | pixel-blind: filename → train label | pixel-blind: source_strip → train majority |
pixel-blind: source_film → train majority |
|---|---|---|---|---|---|---|
| validation | 6,101 | cracks (1,908) |
31.3% | 31.3% | 89.1% | 79.8% |
| test | 2,443 | cracks (765) |
31.3% | 100.0% | 89.7% | 79.7% |
All three probes are fitted on train and never read a pixel — they see only the provenance strings this repo
publishes in metadata. On the honest (crop-disjoint) validation split, knowing which strip a tile came from is
worth 89.1%; on the test split, a filename lookup is worth 100%. Any vision model scoring below those
numbers is losing to a probe that cannot see.
How to use this dataset. Not with the shipped split, unless you are reproducing the paper. Carve film-wise
on metadata.source_film (109 groups) or at least strip-wise on metadata.source_strip (479 groups), and
de-duplicate on metadata.image_sha256 first. Every key you need is in metadata; the split is published exactly
as released because re-cutting it here would freeze one carving choice into the canon — that decision belongs to
the consumer, and this card gives it the numbers to make it with.
Provenance. Unpacked from the 19-part RAR under Dataset_partitioned/ at
github.com/stefyste/RIAWELC (commit-less main, fetched 2026-09-08). The
archive's root folder is literally named DB - Copy — upstream's name, kept in metadata.image_path rather than
tidied away. Every image byte in this repo was checked against that archive: 24,407 / 24,407 identical. The
repository carries no LICENSE file; the README states only that "the dataset is released freely to the research
community", so this card is license: other — confirm terms with the authors before redistribution.
Version history. This repo was first published (2026-07-17) as a raw save_to_disk Arrow dump with a
JSON annot and no card. That layout is preserved in the repo's git history at revision
3874b15e4717c8de76c4b27a2d140bdd1f13ec3e and was replaced, not deleted, by the parquet release documented
here. The images are byte-identical across the two; what changed is the answer column (JSON → plain text),
the metadata (added image_sha256, measured image_wh, the group keys), the dropped all-empty
other_files column, and this card.
Query text — pooled paraphrases (v2)
Every record's query is drawn from common/vision_query_pools.json[102/orig], 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 2.9%).
Template 0 is v1's wording byte for byte (631 records keep it); the pass asserted that on every record before rewriting anything.
Template ↔ gold independence on this build: 24,407 records, 37 templates, worst template p = 0.0874, alpha 2.7e-04, 0 flagged → PASS.
Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 24,407 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: RIAWELC. Upstream license: not stated by the authors (public GitHub release, "released freely to the research community"; the repository carries no LICENSE file — verify before redistribution) (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 102/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.
Converter: forge_model@c0507e2, merged in PR #80 as cbc97aa. 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 (checked by image_sha256 against the org overlap manifest: 0 of 24,407). ⚠⚠ The dataset's OWN official split is fully leaked: every one of the 2,443 test images is a byte-identical copy of a train image — upstream, not from this conversion. See Task, split & the leakage you have to handle above before reporting any number on it.
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 | 24,407 |
| distinct by decoded pixels | 21,893 |
| images carrying more than one record | 2,498 |
| images on both sides of the split | 2,465 |
⚠ This dataset declares a exempt image-identity policy, so the row above is expected to be non-zero: UPSTREAM, documented on this card since v1: the official RIAWELC split ships every one of its 2,443 test images as a byte-identical copy of a train image (same filename under testing/ and training/). Measured on the decoded pixels for v2: 25 validation rows are additionally pixel-identical to a train row (same source film, overlapping strips, re-encoded, invisible to the file hash), plus 44 duplicate groups inside train and 2 inside validation; labels agree throughout. The splits are published exactly as released so the paper's protocol stays reproducible; carve film-wise and de-duplicate on metadata.pixel_sha256 before evaluating Images are still forbidden from crossing the split — and that rule too is exempted here, which is why the last row may be non-zero.
- Downloads last month
- 66