antins-labs/VisSearch-Bench-v1

VisBench-EN

VisBench-EN contains 204 multimodal search questions across the five task categories defined in the VisSearch-Bench paper. All records belong to the test split. Their IDs, categories, domains, image bytes, and within-task image order match VisBench-CN. There are 190 input images across 187 questions. Within Set-wise Visual Search, 17 questions have no initial image; their images value is [].

The release begins with the seven tasks shown in Figure 2 of the paper, followed by the 14 additional unique tasks from the appendix case studies in paper order. The 17 appendix case studies cover 15 distinct tasks, including one already in Figure 2; each task appears only once. The remaining 183 tasks are randomly ordered using the fixed seed 20261103. Each task keeps its original id; the row position is not the task ID. Match records and evaluation results by id.

The questions, reference answers, reference search trajectories, and both evaluation prompts are English translations of the revised Chinese versions. Links, identifiers, facts, conditions, and scoring rules are retained. Input images are unchanged, including any text printed inside them. Literal inscriptions, filenames, and source identifiers may retain their original language when translating them would change what they refer to.

Fields

Field	Type	Description
id	int64	Stable record ID from 1 to 204, matching VisBench-CN
category_name	string	One of the five paper task categories, matching VisBench-CN
domain	string	Subject domain, unchanged from VisBench-CN
question_en	string	English translation of the revised question
answer	string	English translation of the revised reference answer
gold_trajectory	string	English reference search trajectory with source links
images	list[Image]	Input images, with image bytes embedded in Parquet
accuracy_prompt	string	English whole-answer accuracy evaluation prompt
Process Score	string	English evaluation prompt for this record's Process Score

The two evaluation fields contain English prompt text, not precomputed scores. Replace {{question}} with question_en, {{reference_answer}} with answer, and {{model_answer}} with the model response. Read the final score using the output format specified in that record's prompt. Each record keeps its original scoring method and weights.

Task categories

Category	Questions
Set-wise Visual Search	68
Sequential Visual Search	34
Hybrid Wide–Deep Search	34
Competitive Visual Disambiguation	34
Visual-Structure-Guided Search	34

Load the dataset

After uploading this folder's contents to the Hub:

from datasets import load_dataset

ds = load_dataset("antins-labs/VisSearch-Bench-v1", split="test")
sample = ds[0]
question = sample["question_en"]
images = sample["images"]  # A list of PIL images; [] for questions without input images

To read these files locally from this directory:

from datasets import load_dataset

ds = load_dataset("parquet", data_files={"test": "data/test-*.parquet"}, split="test")

The Parquet file includes Hugging Face Image feature metadata and the image bytes. It does not depend on image paths on the export machine. Local-only image_paths, case_id, sources, and hop-related fields are not part of this upload package.