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.

103-discriminate

Discrimination rung on the same 640x640 tiles as 103: three balanced pools (defect-bearing / pseudo-defect-only / blank); the answer is the defect boxes or none, and a pseudo-defect is never reported. 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

29,727 records (test=5918 · train=23809).

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 defect boxes present, one per line as `<class>,[x, y, width, height]` in tile pixels, or `none`. `none` is the correct answer on two thirds of records for two DIFFERENT reasons — nothing there, and something there that is not a weld defect (`伪缺陷`) — and telling those two apart is the task
`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

Why this exists as its own repository

The deposit carries 12,791 伪缺陷 polygons — marks a qualified annotator looked at and recorded as "this looks like a defect and is not". Nothing else in this corpus has that. It teaches discrimination, not detection, and it aims at the failure mode our own PoCs measured: reasoning-mode false positives flooding the answer, and invented detail on records whose evidence is too small to support it.

⚠ It is not a pool of extra negatives for 103. Mixed into the detection set, 伪缺陷 becomes an unlabelled difficulty gradient that shows up only as noise in the main metric. Kept as its own set it produces a number no other dataset here can:

false-positive rate on the 伪缺陷 pool − false-positive rate on the blank pool

A model that reports the same rate on both has learned "is something there", not "is it a defect". The gap is the whole point, and it is measurable only because the pools are separated and balanced.

Records — three pools, balanced, same tiles as `103`

The same 640 × 640 full-scan-resolution tiles as canon 103 (same films, same band, same per-film window; imaging and licence as on 103's card), drawn from three pools of equal size:

pool	records	film-type mix (L / T)
defect-bearing	9,909	4,676 / 5,233
`伪缺陷`-only (no defect)	9,909	4,676 / 5,233
blank	9,909	4,676 / 5,233
total	29,727	equalised across pools by construction (asserted)

The pool size is the smallest of the three (every 伪缺陷-only tile is used); the film-type mix is equalised across pools so that "T-type film ⇒ defect" is not learnable from the pool itself. The blank pool is this repository's own sample (seed 1031), rendered with the same per-film window and renderer as 103, so a tile shared with 103 is the same file. The films 103 drops — one undecodable, 44 duplicates (40 films shipped twice; 4 pairs with conflicting annotations) — are dropped here too.

⚠⚠ The question is "is there a reportable defect" — never "where is the false indication"

伪缺陷 is far smaller than the real defects: at native film resolution its short side is median 17 px with 45.9 % under the corpus's 16 px legibility floor, against 4.4 % for the nine defect classes. That is safe here and only here, because the gold for such a tile is none — a mark the model cannot resolve produces exactly the behaviour the label asks for, and the label is justified by the absence of evidence, which is what is actually visible. Asking the model to point at the false indication would turn this into a find-it task and the floor would apply in full: half the targets would be unresolvable and the label would teach guessing. 伪缺陷 is therefore never in an answer, and metadata.pseudo_legible_at_native per instance lets a later rung filter on measurement.

This reasoning is reasoned, not measured. The experiment that would settle it is an ablation — train with and without sub-floor 伪缺陷, score the false-positive rate — and it is an open item, not an established result.

Split — the same as `103`, film for film

23,809 train / 5,918 test records (19.9 %), by film, identical to canon 103's split.

Lazy-baseline floors

blind predictor	score
always `none`	correct on 66.7 % of records by construction — never report plain accuracy; report the two false-positive rates and their gap
the tile's (width, height)	vacuous by construction — every tile is 640 × 640

Version history

v1 — published 2026-09-13. First publish: 29,727 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-discriminate/verdict_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 2.9%).

Template 0 is the converter's base wording (QUERY in the converter; 842 records carry it); this is the repository's first publish, so there is no earlier wording to reproduce.

Template ↔ gold independence on this build: 29,727 records, 37 templates, worst template p = 0.00532, alpha 2.7e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 29,727 images share one frame size.

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 (defect vs pseudo-defect tiles). 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)

Same tiles as 103 (a subset: 9,909 per pool), same film-wise split. ⚠ Not extra negatives for 103: kept apart so the pseudo-defect false-positive rate is a measurable axis (the metric is the GAP between the pseudo pool and the blank pool). 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	29,727
distinct by decoded pixels	29,727
images carrying more than one record	0
images on both sides of the split	0

Downloads last month: 17

Total file size:

4.73 GB

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