paper-dataset-123456/nips

ManuVision-Bench

ManuVision-Bench: A Benchmark for Image-Grounded Manufacturing Reasoning and Supplier Discovery

Overview

ManuVision-Bench is an evaluation-only candidate-set benchmark for manufacturing reasoning over real-world product images. Each example pairs a product image with a natural-language question and six candidate answer choices. The benchmark evaluates whether multimodal models can connect visual product evidence to manufacturing concepts such as product type, industry sector, manufacturing process, process sequence, and supplier capability.

The benchmark contains 7,661 QA items across 6,443 unique product images. Answers are grounded in external manufacturing and company sources, including D&B Hoovers industry and supplier metadata and Manufactured in North Carolina product and process metadata. Candidate sets are designed with plausible distractors to reduce shortcut solving by easy elimination.

The benchmark is organized as a single dataset with five subsets:

• Q1: Product recognition from product images; multi-answer task.
• Q2: Industry classification from product images; single-answer task.
• Q3: Manufacturing process identification from visual manufacturing cues; multi-answer task.
• Q4: Manufacturing process-chain reasoning; single-answer task.
• Q5: Supplier capability matching from product images and supplier descriptions; single-answer task.

Dataset Structure

ManuVision-Bench/
  README.md
  croissant.json
  Q1/
    benchmark_q1.tsv
    images.zip
  Q2/
    benchmark_q2.tsv
    images.zip
  Q3/
    benchmark_q3.tsv
    images.zip
  Q4/
    benchmark_q4.tsv
    images.zip
  Q5/
    benchmark_q5.tsv
    images.zip

Data Format

Each subset contains a TSV file with the following columns:

• index: question index within the subset
• question: natural-language question grounded in the input image
• image_path: relative path to the corresponding image file
• A to F: candidate answer options
• answer: ground-truth answer label(s)

All images referenced in each TSV are included in the corresponding images.zip archive.

Subset Summary

Subset	Rows	Images	Notes
Q1	839	839	Multi-answer evaluation
Q2	5000	5000	Single-answer evaluation
Q3	616	616	Multi-answer evaluation
Q4	131	131	Single-answer evaluation
Q5	1075	1075	Single-answer evaluation

Task Definition

Given a manufacturing product image, a natural-language question, and six candidate answer choices, the task is to select the correct answer label or labels.

Q1 and Q3 are multi-answer tasks. Q2, Q4, and Q5 are single-answer tasks using a “most likely” formulation. The five subsets evaluate product recognition, industry classification, manufacturing process identification, process-chain reasoning, and supplier capability matching.

Intended Use

ManuVision-Bench is intended for research and evaluation of multimodal models on manufacturing-domain reasoning. It can be used to evaluate visual product recognition, industry and supplier reasoning, manufacturing process understanding, process-chain reasoning, and robustness to plausible distractors.

The benchmark is intended for evaluation only. It is not intended for model training, automated procurement, supplier recommendation deployment, or real-world manufacturing decision-making without human review.

License

The benchmark annotations and metadata are released for non-commercial research and evaluation use under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0), unless otherwise specified.

Images and source-derived company information may be subject to the rights and terms of their original sources. Users are responsible for complying with applicable licenses, terms of use, and attribution requirements for downstream use.

Citation

Citation will be added after the double-blind review process.

Croissant Metadata

This dataset includes a top-level Croissant metadata file:

• croissant.json

The metadata describes ManuVision-Bench as a single benchmark dataset with five subset-specific record sets.