HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents
Paper • 2610.03574 • Published • 43
HyperBrowseComp benchmark access
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
HyperBrowseComp is a held-out evaluation benchmark. To preserve the value of future evaluations, the authors ask users not to use its questions or answers for model training, fine-tuning, or distillation. This is a research practice request, not an additional restriction on the CC BY-SA 4.0 license.
Log in or Sign Up to review the conditions and access this dataset content.
HyperBrowseComp
HyperBrowseComp is a multilingual and multimodal stress test for web-browsing agents. The paper's filtered benchmark contains 423 manually authored and human-validated questions across 13 languages. Each question has a short, publicly verifiable answer, but finding it can require following several clues across obscure webpages, videos, scanned documents, images, or maps. The benchmark tests whether agents can persistently find and connect evidence on the open web.
Project website · Paper · Code
Dataset
The test split is the paper's 423-question release, stored in
data/hyperbrowsecomp_retained_423_encrypted.parquet.
Its columns are id, canary, question, and answer. The question and answer
fields are encrypted with AES-256-GCM; canary is used to derive each field's
nonce. IDs have the form <LANG>_<question id>, such as KO_7388.
License and benchmark integrity
HyperBrowseComp is licensed under CC BY-SA 4.0. To help keep the benchmark useful for evaluation, the authors ask users not to train, fine-tune, or distill models on its questions or answers. This request does not alter the permissions granted by the license.
Load and decrypt
Accept the dataset's access conditions and log in to Hugging Face before
downloading. The test split loads the paper's 423 questions. The decryption
key is included in the final line:
import base64
import hashlib
from datasets import load_dataset
from cryptography.hazmat.primitives.ciphers.aead import AESGCM
ds = load_dataset("afaji/HyperBrowseComp", split="test")
def _nonce(canary, field):
return hashlib.sha256(f"{canary}:{field}".encode()).digest()[:12]
def _open(ciphertext_b64, key, canary, field):
return AESGCM(key).decrypt(_nonce(canary, field), base64.b64decode(ciphertext_b64), None).decode()
def decrypt_row(row, master_key):
key = bytes.fromhex(master_key)
return {
"id": row["id"],
"question": _open(row["question"], key, row["canary"], "question"),
"answer": _open(row["answer"], key, row["canary"], "answer"),
}
decrypt_row(ds[0], "2a8633906fda29da5f696452a1e58f00fcb1edbb3b9c231783f3c1acf80021cb")
- Downloads last month
- 191