BrowseSafe++

Results

Results reported in the BrowseSafe++ paper and community results submitted through the leaderboard Space. Detection tables classify every test page, attacked or control, so a model is rewarded for catching injections and penalised for false alarms on otherwise identical pages.

Accuracy, latency, cost and size

Detection accuracy on the text views against what each model costs to run.

Submit results

Score a detector on the test split and add it to the boards above.

The leaderboard is computed on the 9,175-pair test split. The labels stay private: you run your detector on your own hardware, and the leaderboard Space scores the predictions.

  1. Request access to the dataset on Hugging Face and accept the terms. One request covers the training split and the test inputs: one page of each test pair, the attacked page or its control, licensed for evaluation only.
  2. Run your detector with browsesafe predict from the SDK. It writes one decision per page and view.
  3. Sign in to the leaderboard Space with your Hugging Face account and upload the predictions file.
  4. Receive aggregate scores for each view, with the same metrics as the tables above.

Accounts must be at least 60 days old, and each can submit once per week. Maintainers review every request to publish; approved results appear on this page within about an hour, for every view in which each page has a decision. By default each view shows the top five community results and hides results with an F1 below 50. The submission guide describes the file format and the rules.

shell
pip install "browsesafe-pp[eval] @ git+https://github.com/perplexityai/browsesafe-plus-plus.git"
hf auth login   # log in with the account that has access
browsesafe predict --predict mymod:predict --view html --view markdown \
  --output predictions.jsonl

To develop on the training split first, see Quick start.