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.
103-seam
Weld-seam localisation on whole digitised X-ray films resized to long side 1024 (one or two seams per film; boxes in the resized frame). The one layer of this deposit that survives a full-frame resize. 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
3,632 records (test=724 · train=2908).
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 weld seam, seam,[x, y, width, height] in pixels of the RESIZED frame (long side 1024, top-left origin; the scale rides in metadata.scale). An L film has one seam, a T film two perpendicular ones; the count is part of the answer |
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 — the one layer that survives a full-frame resize
Canon 103 has to be tiled because 90.6 % of its
defect polygons fall under the 16 px legibility floor at long side 1024. The seam does not. Measured
over all seam instances of the deposit, the seam's short side is:
| native | 2048 | 1536 | 1024 | 768 | 512 | |
|---|---|---|---|---|---|---|
| median short side (px) | 707 | 200 | 150 | 100 | 75 | 50 |
| under the 16 px floor | 0.0 % | 0.0 % | 0.0 % | 0.0 % | 0.0 % | 0.1 % |
So the whole film is fed at long side 1024 (min seam short side 24 px on this build) and the answer
is still legible. This is not a defect task: it locates the weld, and 103 inspects it.
Imaging. Digitised X-ray films (SWRD preprint §2, Table 1: Microtek MII5000LC industrial CCD
scanner, 1200 dpi, 16-bit); each film rendered with its own 1–99 percentile grey window
(metadata.window; 280 of the films are 8-bit and pass through unchanged). Every record carries
metadata.source_imaging.
Licence — the permission this rests on
The depositor publishes no terms. Wei recorded on 2026-09-09 a permission from Xuefeng Zhao /
Xinghua Yu (BIT) covering a gated, research-only redistribution of this converted copy, not
transferable; the original correspondence is to be archived with
reports/_notes/grant_records_2026-09.md. Do not redistribute.
Structure is the answer's shape
An L film carries one longitudinal seam; a T film carries two perpendicular ones (a wide
horizontal run and a narrow vertical one). On this build (3,632 films): 2 films have no 焊缝 polygon and are excluded, 1 film is
undecodable as shipped (DJ-RT-20240105-161.tif, zeroed TIFF directory) and 44 films are dropped
as duplicates — the deposit ships 40 films twice under two names (36 pairs with identical
annotations keep the earlier copy; 4 pairs with conflicting annotations lose both) — all counted on
103's card, no frame emitted here. metadata.film_type ∈ {L, T} and metadata.n_seams are on every record
because the count is part of the task: a model that always answers "one horizontal band" is right on
65.8 % of films.
The seam box is the bounding box of the 焊缝 polygon in the resized frame; metadata.scale is the
resize factor and metadata.seams_native the same boxes in film pixels, so nothing is lost. ⚠ The seam
is stored unpadded on purpose: 103's padded band clamps at the film edge, and recovering the seam by
subtracting the pad is not an inverse.
Split — the same as 103, film for film
2,908 train / 724 test films (19.9 %), 0 shared, identical to canon 103's film-wise split (same
drops, film for film) so the two can be trained and evaluated together.
Lazy-baseline floors
| blind predictor | score |
|---|---|
| always "one seam" | right on 65.8 % of films |
| the film's (width, height) → number of seams (class-balanced linear probe) | balanced 0.444 vs 0.500 chance, plain 0.449 vs 0.658 majority, 786 distinct frame sizes — no signal |
Version history
v1 — published 2026-09-13. First publish: 3,632 records, 37 gate-verified query paraphrases
(metadata.query_template; template 0 is the converter's base wording).
Query text — pooled paraphrases (v1)
Every record's query is drawn from common/vision_query_pools.json[103-seam/box], a pool of 37 gate-verified paraphrases of the shipped wording, assigned by a stable hash of the record's image key and recorded as metadata.query_template (37 templates in use, top share 3.3%).
Template 0 is the converter's base wording (QUERY in the converter; 107 records carry it); this is the repository's first publish, so there is no earlier wording to reproduce.
Template ↔ gold independence on this build: 3,632 records, 37 templates, worst template p = 0.0336, alpha 2.7e-04, 0 flagged → PASS.
Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): MAE 0.431 vs 0.345 blind median, 202 distinct frame sizes — no signal.
Built from source by the converter named under Provenance; the pixel-identity guard ran on every image at build time (§8 below).
Provenance
Underlying dataset: SWXD seam-weld radiography (whole films, seam location). Upstream license: No licence is published by the depositor. Use rests on the permission Wei recorded from Xuefeng Zhao / Xinghua Yu (BIT) on 2026-09-09 (reports/_notes/grant_records_2026-09.md): gated research-only redistribution of this converted copy, not transferable — see the Licence section of this card (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 103/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.
Converter: forge_model@aac1823, merged in PR #98 as a4e61f5. 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)
The whole-film companion of 103: the same 3,632 films (same drops: one undecodable, 44 duplicates), same film-wise split, film for film. Locating the seam is not a defect task — 103 inspects what this repository finds. No overlap with any other dataset in this corpus.
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 | 3,632 |
| distinct by decoded pixels | 3,632 |
| images carrying more than one record | 0 |
| images on both sides of the split | 0 |
Cross-family evaluation lock — metadata.eval_lock (stamped 2026-09-20; manifest revision fe6e286912b0, generated 2026-09-08). Every record of this repository, locked or not, carries metadata.eval_lock, computed by forge_model/common/overlap.py::Overlap.stamp_for against common/overlap_manifest.json at that revision — so within this repository the absence of the key cannot occur. Shape: {"locked": bool, "against": [{"repo": …, "split": …}, …], "own_split": …, "manifest_revision": …, "manifest_generated": …}. locked is true when the image is evaluation material anywhere in the corpus; against names every repository and split in which it is (sorted; [] when not locked; it includes the record's own family where that is so); own_split marks a record locked by its own split. The per-record field is the authority — the count here is quoted once, at this revision, and a later manifest may change it: 724 of 3,632 records (724 distinct images) are locked — by column: 0 by the cross-family manifest, 724 by their own split, 0 both ways and counted once; counterparts (records per counterpart; a record can appear under several): none — every lock here is by the record's own split; 724 locked by their own split: test. In words: 724 of the 3,632 records in this repository are evaluation material by their own metadata.split (test: 724) and sit inside the HF split named test / train — under the uniform-split convention the HF split name is a container name, and metadata.split together with metadata.eval_lock carries the truth; a train pool must exclude them. A stamp whose manifest_revision differs from the current manifest is stale, not wrong — recompute it (Overlap.stamp_is_current); a record with no stamp has not been checked against the corpus as it now is. Overlap.partition / assert_train_pool_clean read the field: a train pool built from this repository must exclude every locked record.
- Downloads last month
- 38