jing666888/mro-benchmark

MRO-Bench Dataset

MRO-Bench is a specialized benchmark designed to evaluate next-generation LLMs (equipped with tool-calling, RAG, and complex reasoning) in the domain of Maintenance, Repair, and Operations (MRO) procurement.

1. Architectural Vision

In Industrial MRO scenarios, queries are often unstructured and contain high-density engineering parameters. Unlike general search, a 1mm difference in thread size or a variation in material grade results in a complete system failure. MRO-Bench requires models to act as "industrial selection experts" to identify correct SKUs among highly homogeneous candidates.

2. Taxonomy & Category Descriptions

The benchmark covers 15 major industrial material groups. Each category evaluates specific reasoning capabilities:

Category	Evaluation Focus
electrical	Voltage, current, phase, and industrial safety certifications.
hand_tools	Physical dimensions, drive types, and material alloys.
pumps_pipes_valves	Fluid dynamics, thread standards, and pressure ratings.
security	Access protocols, encryption, and physical protection levels.
ppe	OSHA/ANSI compliance and material resistance ratings.
welding	Output power, duty cycles, and filler metal compatibility.
tape_labels	Adhesion strength, substrate type, and temperature resistance.
power_tools	RPM, torque, and power transmission mechanisms.
pneumatics_hydraulics	Nominal diameters, tolerance limits, and interface types.
factory_automation	Sensor mechanisms and industrial communication protocols.
lighting	Photometric parameters (lumens, color temp) and IP ratings.
office_supplies	Compatibility with existing industrial office hardware.
industrial_control	I/O configurations, relay architectures, and bus protocols.
test_measurement	Metrological ranges, resolution, and accuracy indicators.
chemical_reagents	Purity grades (ACS/GR), concentration, and volume.

3. Dataset Schema

• query_text: Raw user inquiry from real-world business pipelines.
• recall_sku: Candidate pool containing dynamicAttrList.
• dynamicAttrList: Nested technical attributes (e.g., "Thread Size: M10"). This is the core challenge for parametric matching.
• expect_recommended_sku: Ground Truth SKU ID(s).

4. Evaluation Metrics

• Recall@1: The primary metric. The correct SKU must be ranked #1.
• Exact Match (EM): Precise match of the recommended ID sequence.

5. Load Dataset

import os
from datasets import load_dataset
from huggingface_hub import snapshot_download

# Download the repository
repo_id = "jing666888/mro-benchmark"
data_dir = snapshot_download(repo_id=repo_id, repo_type="dataset")

# Load a specific category
dataset = load_dataset(repo_id, "hand_tools", split="validation")

# Example access
sample = dataset[0]
print(f"Query: {sample['query_text']}")
print(f"Target SKU: {sample['expect_recommended_sku']}")