README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
dataset_info:
  features:
    - name: id
      dtype: string
    - name: category
      dtype: string
    - name: sub_category
      dtype: string
    - name: image
      dtype: string
    - name: image_paths
      dtype: string
    - name: encrypted_question
      dtype: string
    - name: encrypted_answer
      dtype: string
    - name: metadata
      dtype: string
    - name: sub_goals
      dtype: string

BrowseComp-V³: A Benchmark Dataset for Multimodal Browsing Agents

A dataset containing 300 samples with encrypted question-answer pairs, images, search trajectories, and sub-goals.

Contents

├── data/
│   ├── train.jsonl          # Main dataset (1.44 MB, 300 samples)
│   └── images/              # Referenced images
├── scripts/
│   ├── decryption_script.py       # Decrypt entire dataset
│   ├── decrypt_batch.py           # Batch decrypt to files
│   ├── encryption_utils.py        # Encryption/decryption utilities
├── metadata/
├── decryption_guide.md      # Decryption instructions
└── README.md

Dataset Format

Fields

Each sample in train.jsonl contains:

• id: Sample identifier (e.g., 001_Culture_Art_L3_p2_t2_m0_r2_w0_c0_g5)
• category: Main category
• sub_category: Sub-category
• image: First image filename
• image_paths: List of image filenames (JSON string)
• encrypted_question: Encrypted question (AES-256-GCM format)
• encrypted_answer: Encrypted answer (AES-256-GCM format)
• metadata: Contains vis_inputs, source, timestamp, level, difficulty, domain, fc_num, and trajectory (JSON string)
• sub_goals: List of sub-goals (JSON string)

Example

{
  "id": "001_Culture_Art_L3_p2_t2_m0_r2_w0_c0_g5",
  "category": "Culture",
  "sub_category": "Art",
  "image": "data/images/001_Culture_Art_1.jpg",
  "image_paths": "[\"data/images/001_Culture_Art_1.jpg\", \"data/images/001_Culture_Art_2.jpg\"]",
  "encrypted_question": "{\"iv\": \"...\", \"ciphertext\": \"...\", \"tag\": \"...\"}",
  "encrypted_answer": "{\"iv\": \"...\", \"ciphertext\": \"...\", \"tag\": \"...\"}",
  "metadata": "{\"vis_inputs\": 2, \"source\": [...], \"level\": 3, \"difficulty\": \"Medium\", \"trajectory\": {...}}",
  "sub_goals": "[{\"sg_id\": 1, \"description\": \"...\", \"key_info\": \"...\", ...}]"
}

Data Structure Notes

• All nested structures (metadata, image_paths, sub_goals, etc.) are stored as JSON strings
• The trajectory field is contained within metadata
• Encrypted fields use base64-encoded iv, ciphertext, and tag

Encryption

Question-answer pairs are encrypted with AES-256-GCM using:

Key derivation: SHA-256 hash of passphrase
Passphrase: A_Visual_Vertical_Verifiable_Benchmark_for_Multimodal_Browsing_Agents

Decryption

Requirements

pip install cryptography

Decrypt Full Dataset

python scripts/decryption_script.py \
  --input data/train.jsonl \
  --key "A_Visual_Vertical_Verifiable_Benchmark_for_Multimodal_Browsing_Agents" \
  --output decrypted.json

Batch Decrypt

echo "A_Visual_Vertical_Verifiable_Benchmark_for_Multimodal_Browsing_Agents" > key.txt

python scripts/decrypt_batch.py \
  --input data/train.jsonl \
  --key-file key.txt \
  --output-dir decrypted_samples/

Using Decryption API

from encryption_utils import derive_key, decrypt_text

key = derive_key("A_Visual_Vertical_Verifiable_Benchmark_for_Multimodal_Browsing_Agents")
encrypted = {"iv": "...", "ciphertext": "...", "tag": "..."}
plaintext = decrypt_text(encrypted, key)

Usage

Load with HuggingFace Datasets

from datasets import load_dataset

ds = load_dataset("path/to/repo")
sample = ds['train'][0]

# Encrypted question and answer are still encrypted
print(sample['encrypted_question'])
print(sample['encrypted_answer'])

Decrypt and Use

import json
import sys
sys.path.insert(0, 'scripts')
from encryption_utils import derive_key, decrypt_text

key = derive_key("A_Visual_Vertical_Verifiable_Benchmark_for_Multimodal_Browsing_Agents")

with open('decrypted.json') as f:
    samples = json.load(f)
    for sample in samples:
        print(f"Q: {sample['question']}")
        print(f"A: {sample['answer']}")

License

CC BY 4.0

---

Last Updated: 2026-02-14