Evaluation benchmark for vision-language models (VLMs) on welding imagery. Four held-out test sets in one repository. It is the evaluation counterpart of jujeongho/Weld-Instruct-v2; nothing in Weld-Bench may be used for training.

v2 keeps the rows and images of Weld-Bench v1 and changes the prompts and answers so that every subset follows the prompt of a public evaluation benchmark:

multiple choice follows the official MMMU format (the default multiple-choice instruction of lmms-eval);
grounding and segmentation follow the HumanRef task of lmms-eval: every class is asked for every image, an answer may hold zero, one or several objects, and the scoring is HumanRef's DensityF1 and rejection accuracy.

Subset	Task	Rows	Image modality	Metric
`RIAWELC`	4-way defect classification	2,443	X-ray radiograph patches, 227×227 grayscale	accuracy
`LoHi-WELD-hweld`	defect grounding, one row per (image, class)	820 (205 images × 4 classes)	robotic MAG weld-bead crops (high-resolution subset), RGB	DensityF1@0.5:0.95 · rejection accuracy
`RefWeld`	component grounding (box) and segmentation (polygon), one row per (image, class, task)	1,188 (85 images × 7 classes × box/poly, see below)	screenshots of welding videos	DensityF1@0.5:0.95 · rejection accuracy, per task
`Khvatskii`	weld acceptability judgement per use context	459 (153 images × 3 contexts)	field / web photos of welds	precision · recall · F1

All setting: zero-shot. Prompts name the classes but give no class definitions.

Access and loading

The repository is public but gated (manual approval). After approval:

from datasets import load_dataset
riawelc = load_dataset("jujeongho/Weld-Bench-v2", "RIAWELC", split="test")
lohi    = load_dataset("jujeongho/Weld-Bench-v2", "LoHi-WELD-hweld", split="test")
refweld = load_dataset("jujeongho/Weld-Bench-v2", "RefWeld", split="test")
khv     = load_dataset("jujeongho/Weld-Bench-v2", "Khvatskii", split="test")

Schema. RIAWELC, LoHi-WELD-hweld and RefWeld have sample_id, image (PIL image, original pixels, never resized), question and answer (strings). Khvatskii has sample_id, image, system, question1, question2, answer and narrative (the expert's rationale; not scored).

Common protocol

Prompt: one user turn with the image followed by the question string, verbatim. No system prompt, no few-shot examples, no extra format hints.
Image: send the image column as is. If an engine must downscale, cap the longer side at 1536 px (only Khvatskii images exceed it) and use the same cap for every model.
Decoding: greedy (temperature 0) unless the subset says otherwise; reasoning / "thinking" modes off for every model; max_tokens ≥ 2048. If a model still emits a reasoning block, only the text after the last </think> is scored.
Coordinates in the answers: integers in a 0–999 frame normalized to the image (x = column / width × 1000, y = row / height × 1000; x1, y1 = top-left, x2, y2 = bottom-right). The prompts do not state a frame (as in HumanRef); see the coordinate rule in the grounding section.
Report the model size, decoding settings and the number of truncated replies (finish_reason == "length") next to every score.

RIAWELC — radiograph defect classification

Source: Totino, Spagnolo, Perri, RIAWELC: A Novel Dataset of Radiographic Images for Automatic Weld Defects Classification (ICMECE 2022) and Perri et al., Welding Defects Classification Through a Convolutional Neural Network (Manufacturing Letters 2023); data from github.com/stefyste/RIAWELC ("released freely"; cite both papers). We use the authors' official testing folder (2,443 patches) unchanged.

Rows: classes crack 765, porosity 632, lack of penetration 446, no defect 600. Option order is shuffled per row (same order as v1); answers are balanced across letters.

Question (MMMU format):

This is a patch cropped from an industrial weld radiograph (X-ray). Classify it.
Options:
(A) lack of penetration
(B) no defect
(C) crack
(D) porosity
Answer with the option's letter from the given choices directly.

Answer: the letter only, e.g. C.
Scoring: the prediction is the first (X) in the reply; else a leading bare letter (C, C., **C**); else the single option name mentioned in the reply (two or more names = no prediction). Metric = accuracy over 2,443 rows. Chance is 25%.

Grounding and segmentation: LoHi-WELD-hweld and RefWeld

Both subsets use the prompt and the metric of the HumanRef task in lmms-eval (lmms_eval/tasks/humanref/utils.py, a port of IDEA-Research/HumanRef metric/recall_precision_densityf1.py). Every class of the subset is asked for every image, so some rows have no target.

Question:
- box rows: Locate <class>, output its bbox coordinates using JSON format.
- polygon rows (RefWeld only): Locate <class>, output its polygon coordinates using JSON format.
Answer: a JSON list with one entry per object of that class, [{"bbox_2d": [x1, y1, x2, y2], "label": "<class>"}, ...] or [{"polygon_2d": [[x1, y1], ..., [x8, y8]], "label": "<class>"}, ...], and [] when the class is not in the image. Boxes are listed by the x + y of their top-left corners, smallest first; polygons have 8 points in clockwise order starting from the vertex with the smallest x + y. Order does not affect the score.
Parsing (as HumanRef): take the text after the last </think>; an empty reply, [], {}, null or none means no prediction. Otherwise parse JSON (first fenced block, the whole reply, or the outermost [...] / {...}) and collect every box under the keys bbox_2d, bbox, box, coordinates, bounding_box or any bare 4-number list; if none is found, collect every bracketed 4-number tuple with a regex. Boxes with x2 <= x1 or y2 <= y1 are dropped. Polygons are read from polygon_2d, polygon, points or any bare list of ≥ 3 [x, y] points. The label field is not used: the class is the one in the question.
Coordinate rule (as HumanRef): the whole reply is read in each of four conventions — absolute pixels, a 0–1000 grid, 0–1 (only if every value is ≤ 1) and Qwen2.5-VL resized pixels (smart_resize, factor 28) — and the reading with the best DensityF1@0.5:0.95 is kept, so models are not penalised for their native format. The answer is mapped to pixels with v / 1000 × size.
Scoring, rows with a target: for each IoU threshold t ∈ {0.50, 0.55, …, 0.95}, every ground-truth object takes the unused prediction with the highest IoU ≥ t (greedy one-to-one matching); recall = matches / #GT, precision = matches / #predictions, and DensityF1 = F1 × min(1, #GT / #predictions), which penalises predicting more objects than there are. The row's score is the mean over the 10 thresholds. DensityF1@0.5:0.95 is the mean over all rows with a target. Box rows use box IoU in pixels; polygon rows use mask IoU: both polygons are rasterized as binary masks at the image size (pixels inside = 1), IoU = intersection pixels / union pixels.
Scoring, rows without a target: the row is correct if the reply contains no prediction. Rejection accuracy is the share of correct rows among rows without a target.
Report DensityF1@0.5:0.95 and rejection accuracy for each subset (and for RefWeld box and polygon separately).

LoHi-WELD-hweld — weld-bead defect grounding

Source: Block, da Silva, Lazzaretti, Minetto, LoHi-WELD: A Novel Industrial Dataset for Weld Defect Detection and Classification, a Deep Learning Study, and Future Perspectives (IEEE Access 2024); data from github.com/SylvioBlock/LoHi-Weld (free for research and commercial use with citation). Rows are the authors' official hold-out test crops of the high-resolution subset (hweld); the low-resolution subset is excluded because its ~10 px boxes are beyond the localization precision of current VLMs.

Rows: 205 crops (~785×195 px) × 4 classes = 820 rows; sample_id = lohi_high_<crop>_<class>. 609 rows have a target (1,928 boxes: pore 86, deposit 390, discontinuity 596, stain 856), 211 rows have none.
Classes: pore, deposit, discontinuity, stain. For readers only (not shown to the model): pore = small fill gap from an air bubble; deposit = excess welding material along the bead; discontinuity = partial absence of weld material along the bead; stain = slight alteration of the weld surface.
Answer coordinates: from the authors' YOLO labels, round(v × 1000) clipped to 0–999.

RefWeld — welding-component grounding and segmentation

Own data. 85 screenshots of publicly available welding videos (torch-side camera views), annotated with polygons by four people in CVAT, reviewed and corrected by one reviewer, near-duplicate frames removed. Every class has at most one instance per image.

Classes: wire, nozzle, pool, bead, arc, groove, seam. For readers only: wire = filler wire protruding from the nozzle; nozzle = gas nozzle / contact tip; pool = molten pool under the arc; bead = solidified weld metal; seam = joint line before welding; groove = joint gap; arc = electric discharge at the wire tip.
Rows: sample_id = refweld_img<001–085>_<class>_<box|poly>.
- box: 595 = 85 × 7 (396 with a target, 199 without);
- poly: 593 (394 with a target, 199 without) — two pool instances whose 8-point polygon self-intersected have no polygon answer, so their poly rows are left out.
Polygon answers: the annotated polygon canonicalized to 8 points (Visvalingam–Whyatt removal when there are more vertices, longest-edge midpoint insertion when fewer), clockwise on screen, starting at the vertex with the smallest x + y. The model may output any number of points.

Khvatskii — weld acceptability per use context

Unchanged from v1 (the paper's own protocol).

Source: Khvatskii, Lee, Angst, Gibbs, Landers, Chawla, Do Multimodal Large Language Models Understand Welding? (Information Fusion 120, 2025; arXiv 2503.16537); data and prompts from github.com/x0wllaar/Do-LLMs-Understand-Welding.

Rows: 459 = 153 images × 3 use contexts (RV & Marine, Aeronautical, Farming). sample_id = khv_<real|online>_<guid8>_<rvmarine|aeronautical|farming>; real = 95 photos taken by welding instructors, online = 58 web photos.
Conversation (per row, independent): system → user [image + question1] → model turn 1 (≤ 5 sentences of plain text) → user [question2] → model turn 2 (a JSON object {"acceptable": true|false}).
Sampling: run each row 3 times with seeds 0, 1, 2 at temperature 0.7, top_p 0.8; the row's prediction is the majority vote of the three turn-2 answers. Turn 2 is parsed with "acceptable"\s*:\s*(true|false), falling back to a bare true / false; an unparsable turn 2 counts as false.
Answer: {"acceptable": true} or {"acceptable": false}.
Scoring: positive class = acceptable. Report precision, recall and F1 for each context × source cell and the micro F1 over all contexts for real (285 judgements) and online (174), which are the headline numbers.
Reference (same protocol, from the paper): micro F1 real / online — GPT-4o 0.536 / 0.675, LLaVA-1.6-Mistral-7B 0.210 / 0.481.

Changes from v1

Subset	v1	v2
RIAWELC	options in one line, answer `(C) crack`	MMMU format, answer `C`
LoHi-WELD	one question per image (all classes, labelled boxes), COCO mAP@0.5	renamed `LoHi-WELD-hweld`; one question per (image, class) in HumanRef format, DensityF1 + rejection accuracy
RefWeld	one question per annotated (image, class), MAttNet accuracy / Precision@0.5	every class for every image, HumanRef format, DensityF1 + rejection accuracy (mask IoU for polygons)
all coordinates	0–1000	0–999
Khvatskii	—	unchanged

Citation

Please cite the source datasets (RIAWELC [both papers], LoHi-WELD, Khvatskii et al.) and, for the prompt and metric, MMMU (Yue et al., CVPR 2024), HumanRef / RexSeek (Jiang et al., 2025) and lmms-eval.

Downloads last month: 234

Total file size:

700 MB