IndustryBench-MIPU: Benchmarking Multi-Image Attribute Value Extraction for Industrial Products

Multi-Image Industrial Product Understanding Benchmark — evaluating MLLMs on structured attribute extraction from real-world industrial product images.

Industrial product specifications are scattered across multiple heterogeneous images — specification tables, nameplates, technical drawings. IndustryBench-MIPU tests whether MLLMs can reliably recover them through four challenges: text recognition, visual reasoning, domain knowledge, and cross-image evidence integration.

---

Main Results

Model	Multi-P	Multi-R	Multi-F1
Gemini 3.1 Pro	93.8	49.9	65.1
Qwen 3.5-397B-A17B	88.2	48.6	62.7
GPT-5.4	86.3	46.6	60.5
Qwen 3.5 Plus	88.1	45.4	59.9
Claude Opus 4.6	88.2	42.3	57.2

Key finding: Models achieve high precision (86–94%) but the best recovers only 49.9% of product-level attributes. Multi-image completeness is the core bottleneck.

---

Dataset Summary

IndustryBench-MIPU is the first large-scale benchmark for multi-image industrial product understanding. Given product images and a product-specific attribute schema, the task is to extract all property-value pairs.

Products	Images	Top-level Categories	Leaf Categories	Property Names	Product-level Annotations
4,481	26,310	18	2,301	4,554	95,024

The multi-image benchmark covers 4,481 products across 26,310 images. The single-image benchmark provides 6,000 image-level instances (44,111 property-value pairs), a stratified subset selected from the full single-image pool. The benchmark is constructed through multi-model consensus (5 frontier MLLMs) and three-tier quality assurance (frontier model audit + gold-standard cross-check + human verification).

---

Dataset Structure

Files

File	Description	Size
data/multi_image_level.jsonl	Multi-image product-level benchmark (4,481 products)	~9 MB
data/single_image_level.jsonl	Single-image level benchmark (6,000 rows, a stratified subset of the full single-image pool)	~5.6 MB
data/metadata.jsonl	Image-level metadata (mirrors single_image_level.jsonl) with file_name for the Hugging Face viewer (6,000 rows)	~5.9 MB
data/images/	Product images (26,310 JPEG files, sharded into subdirectories)	~5.3 GB

Multi-Image Level (multi_image_level.jsonl)

Each line is one product with all its images and ground-truth attributes:

{
  "item_id": "560324848370",
  "title": "日亚铝基板 NICHIA ...",
  "cate1_name": "电子元器件",
  "cate_name": "铝基板",
  "main_entity": "NICHIA铝基板",
  "cpv_schema": ["颜色", "品牌", "型号", "..."],
  "image_count": 7,
  "images": [{"record_id": "560324848370#main", "image_source": "main_image", "image_index": 0, "image_path": "images/part-000/560324848370_main.jpg"}],
  "cpv_results": {"颜色": ["白色"], "品牌": ["NICHIA"], "...": ["..."]}
}

Single-Image Level (single_image_level.jsonl)

A stratified subset of 6,000 images selected from the full single-image pool. Each line is one image with attributes visible in that specific image:

{
  "record_id": "589158697373#detail_3",
  "image_path": "images/part-001/589158697373_detail_3.jpg",
  "cpv_schema": ["含量", "纯度", "..."],
  "cpv_results": {"含量": ["98%"]}
}

Field Descriptions

Field	Type	Description
item_id	string	Unique product identifier
title	string	Product title (Chinese)
cate1_name	string	Top-level category (18 total)
cate_name	string	Leaf-level category
main_entity	string	Target product entity (anchors extraction)
cpv_schema	list	Valid property names for this product
image_count	int	Number of images (multi-image only)
images	list	Image metadata with paths (multi-image only)
record_id	string	{item_id}#{main|detail}_{index} (single-image only)
image_path	string	Relative path to image file
cpv_results	dict	Ground-truth mapping: property_name → list of property values

---

Usage

Images are provided as regular files under data/images/, split into part-* subdirectories to satisfy Hugging Face repository limits. The data/metadata.jsonl file mirrors the single-image benchmark and adds a file_name field for the Hugging Face dataset viewer.

Run evaluation

See the GitHub repository for the full evaluation pipeline (Extract → Eval → Aggregate).

pip install -r code/requirements.txt

cd code
python run_multi_extract.py \
    --input ../data/multi_image_level.jsonl \
    --output results/extract.jsonl \
    --provider openai --model MODEL \
    --api-key $API_KEY --api-base $API_BASE_URL

python run_eval.py \
    --input results/extract.jsonl \
    --output results/eval.jsonl \
    --provider openai --model MODEL \
    --api-key $API_KEY --api-base $API_BASE_URL

python aggregate_eval.py \
    --input results/eval.jsonl \
    --bench ../data/multi_image_level.jsonl \
    --extract results/extract.jsonl

---

Construction Pipeline

1. Stratified sampling — Random sample per leaf category. All images are sourced from Alibaba's own industrial e-commerce platform.
2. Multi-model annotation — 5 MLLMs independently annotate via a 3-stage pipeline (entity recognition → image filtering → per-image extraction)
3. Union + deduplication — Take union of all outputs, normalize surface forms, semantically deduplicate
4. Three-tier QA — Frontier model audit (23.9% filtered) → Gold-standard cross-check → Human verification (96.7% pass)

---

Citation

@article{industrybench-mipu,
  title={IndustryBench-MIPU: Benchmarking Multi-Image Attribute Value Extraction for Industrial Products},
  author={Multimodal and Industrial AI Team, Alibaba},
  journal={arXiv preprint arXiv:2606.14383},
  year={2026},
  url={https://arxiv.org/abs/2606.14383}
}