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.

108-bead-width

Melt-pool width regression in wire-arc additive manufacturing (continuous target, millimetres). 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

2,549 records (test=1231 · train=1318).

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 (decimal mm)	the answer — for this dataset: the bead width in millimetres to one decimal place, e.g. `9.5`. This is a CONTINUOUS quantity and must be graded numerically — MAE, or hit-rate within a stated tolerance — never by string match. It is upstream's own per-frame measurement taken verbatim from `WAMVID_metadata.csv`; nothing here is derived or interpolated
`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

⚠⚠ Licence — read this before using or redistributing

WAAM-ViD is deposited at Cranfield CERES under CC BY-NC-ND 4.0 — Attribution, NonCommercial, NoDerivatives — read from the repository's own metadata record, not from the paper.

ND means derivative works may not be distributed, and this repo is a derivative: frames extracted from the videos, one copy selected per frame, annotations re-serialised into a different schema.

Redistributed under written permission from the WAAM-ViD authors (Cranfield University), granted 2026-09-08, for gated research-only, non-commercial use. That permission covers this repository's distribution of the derived data. The upstream deposit itself remains CC BY-NC-ND 4.0, and the permission granted to this project does not extend to onward redistribution by anyone else — if you are granted access here you may use the data for non-commercial research, but you may not re-publish it. license: other in the YAML header is the corpus convention; the sentences above are the operative terms.

Cite: Asif, Suder, Kim, Kamerkar, Chiu & Abdi, "WAAM-ViD: towards universal vision-based monitoring for wire arc additive manufacturing", Frontiers in Manufacturing Technology, 2025; data at DOI 10.57996/cran.ceres-2763.

Task & answer

Melt-pool width regression. query asks for the width of the deposited bead at that moment in millimetres; annot is the number to one decimal place, e.g. 9.5.

⚠ This is a continuous target. Grade it numerically — MAE, or hit-rate within a stated tolerance — never by string match. 9.4 against a gold of 9.5 is a 0.1 mm error, not a wrong answer.

The value is upstream's own per-frame measurement, taken verbatim from WAMVID_metadata.csv (widths measured after deposition with a Micro-Epsilon optoNCDT 2300-100 laser profile sensor and aligned to the video frame by the authors). Nothing here is derived, interpolated or rounded beyond upstream's own precision.

⚠ 0.0 is never an answer here. Upstream marks an unmeasured frame nan, and no upstream video contains a single literal 0.0. A frame with no measurement is simply absent from this repo — it is not published as "0.0 mm". (The 2026-07-17 release of 108 did exactly that on 2,420 rows.)

Records, and what was excluded

	frames
annotated source frames in `AI4Manufacturing/108`	4,674
− width not measured upstream (`nan`; the pre-arc lead-in)	−1,186
− video 1 (its annotation frame numbering is inconsistent with its own video)	−502
− source-frame selection ambiguous (a crop/zoom would change mm-per-pixel)	−437
published	2,549 (train 1,318 / test 1,231)

Width range 5.5–13.5 mm; median 9.1.

Why video 1 is out. Its video has 623 frames and its width series has 623 entries — they agree exactly — but upstream's annotation filenames reach frame 747, and 104 of its annotated frames name indices that do not exist in the video. Its frames remain in 108 for localisation, where a polygon still matches its own image; there is no defensible way to attach a measurement to them.

Why ambiguous frames are out. Each training frame ships as three Roboflow augmentations and the un-augmented one is identified by matching against the decoded source video. Where that match is not confident, a crop/zoom variant might have been selected — which changes the millimetres-per-pixel relationship and so corrupts a metric target, even though the label itself is per-frame and identical across copies. The confidence, its calibration table, and the selection_ratio >= 3.0 route to a stricter subset are documented on AI4Manufacturing/108; every record here carries selection_ratio, measures_agree, both distances and the candidate count.

⚠ Video 3 is thin here on purpose. 82.5% of its frames (113 of 137) are selection-ambiguous — by far the worst of the 24 videos — so only 24 of its frames survive into this task. Treat any per-video result for video 3 as unsupported.

Two cameras, two frame sizes

The release mixes 1280×1024 (4,253 frames; videos 1, 3, 5, 11, 15, 18, 20, 21, 22, 24, 25, 30, 31, 35, 36, 37, 41, 45, 53, 69) and 1440×1080 (421 frames; videos 81, 90, 100, 110). The paper names two cameras — a Xiris XVC-1000 passive HDR and a CAVITAR C400-H active laser-illuminated unit — but the release does not say which camera shot which video, so no model name is asserted per record. What is recorded is the fact: metadata.image_wh, plus upstream's own per-video capture record in metadata.capture (camera matrix, distortion coefficients, rotation matrix, translation vector, external lighting flag, arc current, travel speed). Anything computing mm-per-pixel or a legibility floor needs those intrinsics, and they travel with every row rather than living only in a CSV.

Burned-in text — measured as a shortcut for the width, and it is not one

Every 1280 × 1024 frame (4,253 of 4,674) carries a camera timestamp top-left and a "© 2025 Xiris Automation Inc" watermark bottom-left; the 421 frames at 1440 × 1080 carry neither. A linear probe on the timestamp corner alone, fitted on train and scored on test, gives MAE 1.713 mm against a 0.970 mm blind median — worse than guessing, because the timestamp names the session and the test videos are unseen; the watermark strip gives 0.969 (nothing); the frame's (width, height) alone gives 0.960, a 1% edge from the sensor format. A same-size patch from the image centre gives 0.753 – 0.811. The width is in the bead, not in the corner (tools/burned_in_probe.py; images published uncropped). Instrument: closed-form ridge, λ = 10 on the 576-value corner patch; the (width, height) row is common/lazy_floors.py (λ = 1, two features) — stated because a number without its setting has no provenance.

Lazy-baseline floors — and one that behaves the opposite way round

On the test split, MAE in millimetres and hit-rate within ±0.5 mm:

fixed prediction	MAE	within ±0.5 mm
the train median (9.2 mm), answered constantly	0.970 mm	40.9%
the train frame-number prior (median per 25-frame bin)	1.141 mm	31.6%

Report against the better of these — 0.970 mm — not against zero.

⚠ The temporal prior is worse than a constant, which is not what one would guess. The bead does widen as deposition proceeds, so frame index looks like it should predict width. It does — within a video. But train and test are different videos, and the frame→width curve does not transfer across them, so a prior fitted on train actively hurts on test. Any temporal baseline has to be fitted within a video, which this split deliberately does not allow. This is worth knowing before someone reports a "strong" temporal result obtained by fitting inside the test videos.

Split. The release's own Train/Test division, which is video-disjoint: no video contributes to both sides. metadata.video_id is the group key. Any re-carve must be video-wise — consecutive frames of one deposition are near-duplicates, so a frame-level random split is meaningless.

Version history

This repo is new. The width task previously existed only inside AI4Manufacturing/108's 2026-07-17 release, where it was mixed into the same annot as the segmentation answer, keyed by two different and undocumented alignment rules, and published 0.0 for unmeasured frames. See that repo's card for what was replaced.

Query text — pooled paraphrases (v2)

Every record's query is drawn from common/vision_query_pools.json[108-bead-width/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 3.3%).

Template 0 is v1's wording byte for byte (66 records keep it); the pass asserted that on every record before rewriting anything.

Template ↔ gold independence on this build: 2,549 records, 37 templates, worst template p = 0.0254, alpha 2.7e-04, 0 flagged → PASS.

Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): MAE 0.960 vs 0.970 blind median, 2 distinct frame sizes — no signal.

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: WAAM-ViD. Upstream license: CC BY-NC-ND 4.0 upstream (Cranfield CERES; DOI 10.57996/cran.ceres-2763). Redistributed under written permission from the WAAM-ViD authors (Cranfield University), granted 2026-09-08, for gated research-only, non-commercial use — see the licence note below (this card is license: other; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion code under 108/ (with publish/push_to_hf.py) in AI4Manufacturing/forge_model.

Converter: forge_model@19ec06e, merged in PR #84 as 37a13ca. 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)

A 2,549-frame SUBSET of AI4Manufacturing/108 (4,674 frames) — same images, same image_sha256, so keep the two on the same side of any split. The splits are video-disjoint (15 train videos / 8 test); metadata.video_id is the group key.

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	2,549
distinct by decoded pixels	2,549
images carrying more than one record	0
images on both sides of the split	0

Downloads last month: 19

Total file size:

161 MB