Training data for adapting a vision-language model to the welding domain in three stages, following the curriculum of LLaVA-Med (concept alignment → instruction tuning → fine-tuning on downstream datasets). The evaluation counterpart is jujeongho/Weld-Bench-v2; no Weld-Bench test image is used here.

Config	Stage	Rows	Content
`stage1-alignment`	1. concept alignment	72,003	web images of welding (Re-LAION links only) + new captions, English / Korean 1:1
`stage2-instruction`	2. instruction tuning	69,835 conversations (411,149 turns)	multi-turn QA on the stage-1 images (Re-LAION links only), English / Korean
`stage3-downstream`	3. downstream fine-tuning	45,782	training splits of the Weld-Bench sources (RIAWELC, LoHi-WELD) + NC AI in-house grounding data, in the Weld-Bench v2 prompt format

stage1-alignment

Images are not redistributed. Each row points to an image of Re-LAION-2B-en (research-safe) and carries our caption.

Columns: key (Re-LAION key), url, hash (Re-LAION 64-bit id), md5 (md5 of the original image bytes, to check a re-download), lang (en / ko), instruction, caption.

Construction (domain name = "welding"; every step is domain-agnostic):

Keywords: the titles of all 214 pages of Wikipedia's Category:Welding.
Alt-text search over the 2.1 B rows of Re-LAION-2B-en research-safe: lower-case, NFKD with combining marks removed, word-sequence match, plural (s / es) allowed → 777,614 rows; 289,950 images downloaded.
Near-duplicate removal inside the set: Yokoo ISC21 descriptor (isc_ft_v107), cosine > 0.75 groups; per group the image with the highest CLIP ViT-L/14 image–alt-text similarity is kept → 137,958.
Evaluation contamination removal: ISC cosine > 0.604169 (the DataComp threshold) against every image of the evaluation benchmarks (11 general benchmarks and Weld-Bench) → 137,864.
Relevance: Qwen3.5-397B-A17B is shown the image and the alt text as a hint and asked whether the image is directly related to the domain (Yes / No, decided by P("Yes") > P("No") of the first token) → 79,698.
Captioning: Qwen3.5-397B-A17B with the alt text as a hint; only details that can be read or clearly seen in the image are kept; no length limit; language assigned per image (sha1 of the key) for an English / Korean 1:1 split. The prompt is in English; for Korean images one line asking for a Korean description is added.
Verification: gemma-4-31B-it is shown the image and the caption and asked whether the caption is a correct description (Yes / No, decided by P("Yes") > P("No") of the first token); captions truncated at 1,024 tokens are dropped → 72,003 (English 35,177 / Korean 36,826).
Instruction: a LLaVA-Med stage-1 instruction (and its Korean equivalent) — brief for captions of ≤ 2 sentences (7,067), detailed otherwise.

Use of Re-LAION is subject to its terms (released for research; LAION advises against industrial use).

stage2-instruction

Images are not redistributed (same Re-LAION images as stage1-alignment). Each row is one image with a multi-turn conversation about it.

Columns: key, url, hash, md5, lang (en / ko), conversations ([{"role": "user", "content": …}, {"role": "assistant", "content": …}, …]).

Construction (input = the 72,003 verified image–caption pairs of stage1-alignment; the conversation language follows the caption language):

Questions: Qwen3.5-397B-A17B sees the image and the verified caption and writes the questions a person seeing the image for the first time would ask in sequence. The caption is used only to judge what can be confirmed from the image; questions may not require information that is only in the caption. No task list, examples or domain name are given; the model decides the kinds and the number of questions.
Answers: the same model sees only the image and the questions (no caption) and answers them in order; it says when something cannot be determined from the image and corrects false premises.
Turn verification: gemma-4-31B-it sees the image, the earlier turns and the current question–answer pair and is asked whether the answer is correct for the image and consistent with the earlier conversation (Yes / No, decided by P("Yes") > P("No") of the first token).
Assembly: each conversation is cut at the first turn judged No (later turns may refer to earlier ones); conversations with no turn left are dropped. No separate language filter is applied, so answers may quote text written in the image in another language (e.g. an English product name in a Korean conversation). 479,144 generated turns → 411,149 kept; 69,835 conversations (English 34,265 / Korean 35,570).

All prompts are in English; for Korean images one line asking for Korean output is added.

stage3-downstream

Columns: source (RIAWELC, LoHi-WELD or NC AI in-house), image, conversations ([{"role": "user", "content": …}, {"role": "assistant", "content": …}]).

All questions use the Weld-Bench v2 prompt format (see that dataset card): multiple choice in the MMMU format with a letter answer, and grounding / segmentation with the HumanRef question Locate <class>, output its bbox|polygon coordinates using JSON format. and a Qwen-style JSON answer, a list of {"bbox_2d" | "polygon_2d": …, "label": <class>} in a 0–999 normalized frame (polygons have 8 points), or [] when the class is absent.

Source	Rows	Notes
RIAWELC	13,420	the authors' official `training` folder minus the 2,443 files that are byte-identical to the official `testing` folder (= the Weld-Bench test set); the `validation` folder is not included; option order shuffled per row
LoHi-WELD	3,268	high-resolution subset: 817 crops (the official hold-out test excluded) × 4 classes; coordinates from the authors' YOLO labels
NC AI in-house	29,094	box (4,885) and polygon (24,209) grounding of 7 welding components: arc, bead, groove, nozzle, pool, seam, wire

Grounding rows ask one class per question; every class of the source is asked for every image, and a "class absent" question ([]) is kept only when absence is certain from the annotation. Absent questions are sampled down to at most half the number of present questions (stratified by class), close to the share in Weld-Bench (26–33%).

Overlap check with Weld-Bench (2026-11-05). Every training image was compared with every Weld-Bench v2 test image by file md5 and by ISC21 descriptor similarity. The authors' RIAWELC training folder turned out to contain all 2,443 testing images byte for byte; these were removed in this revision (an earlier upload of this config included them). After removal no training image is identical to a Weld-Bench image. ISC similarity above the DataComp threshold remains between crops of the same radiographs (RIAWELC) or the same specimens (LoHi-WELD), which is inherent to the official splits.

Sources and citation

Re-LAION-2B-en research-safe (LAION).
RIAWELC: Totino, Spagnolo, Perri (ICMECE 2022); Perri et al. (Manufacturing Letters 2023) — "released freely", cite both papers.
LoHi-WELD: Block, da Silva, Lazzaretti, Minetto (IEEE Access 2024) — free for research and commercial use with citation.
NC AI in-house data.

Downloads last month: 6,087

Total file size:

3.54 GB