s4um1l/accounting_grounding_v1

Accounting UI Grounding Benchmark

A visual grounding benchmark for professional accounting software interfaces.

Motivation

Existing GUI benchmarks (ScreenSpot, Mind2Web) primarily evaluate consumer applications. Enterprise software presents different challenges: dense information layouts, small interactive elements, and domain-specific terminology. This benchmark measures how well visual grounding models transfer to this domain.

Results

Model	ScreenSpot-Pro	This Benchmark	Delta
ShowUI-2B	75.1%	0.15%	-74.95pp
GroundNext-7B	50.2%	13.28%	-36.92pp

Dataset

Metric	Value
Platforms	Odoo, QuickBooks Online, ERPNext, Dolibarr
Screenshots	65
Grounding Targets	595
Resolution	1920×1080

Platform Breakdown

Platform	Accuracy	Targets	UI Framework
Odoo	45.98%	87	OWL
QuickBooks Online	9.84%	193	React
Dolibarr	7.26%	234	PHP
ERPNext	3.70%	81	Frappe

Failure Analysis

Category	Percentage	Description
Semantic Aliasing	42.8%	Similar labels (Save/Submit, Bill/Invoice)
Icon Scale	33.7%	Small elements (25-35px)
Coordinate Regression	23.4%	Correct element, imprecise coordinates

Data Format

Each sample includes:

• PNG screenshot (1920×1080)
• JSON annotation with bounding boxes, instruction variants, and element metadata

platform/
├── sample.png
└── sample.json

Usage

from datasets import load_dataset

dataset = load_dataset("s4um1l/accounting_grounding_v1")

Citation

@misc{srivastava2026accounting,
  title={Accounting UI Grounding Benchmark},
  author={Srivastava, Saumil},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/s4um1l/accounting_grounding_v1}
}

License

CC BY 4.0

References

• GroundCUA — ServiceNow Research
• ShowUI — ShowLab
• ScreenSpot — GUI Grounding Benchmark