Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LiveWeb Samples

Three browser tasks and a login task on live websites, built by HUD. Each task starts an agent on a real site and grades its answer against a rubric.

Taskset: https://www.hud.ai/tasksets/1fc0f549-5830-4061-bc3a-25ad502ce723/tasks

Tasks

Task What the agent has to do Grading
ladybird-spot-num Turn a 3D ladybird on Sketchfab and count the spots on its shell, head excluded One integer. Each spot off costs 20 points, and a range or several totals scores 0.
diamond-rose-seats Find the kitchen photos on diamondrosesanctuary.com and count the chairs at the table Exact count
epa-air-quality Use EPA's AirNow map to find air quality data near zip 35173 on April 19, 2024 Site, site ID, pollutant, and daily AQI, weighted 20/20/30/30
costco-login Sign in to an existing Costco account, entering the email and password through a tool that keeps them out of tool output Reaching sign-in, entering the credentials without exposing the password, a confirmed signed-in session, and an accurate report with no account changes, 25% each

Answers and full rubrics are in tasks/samples.yaml.

costco-login runs on the rfp-liveweb environment and needs your own Costco account, saved in HUD as a credential profile named costco. Without one, it fails when you run with --all.

Results

Six runs per model per task, September 2026. Scores are mean rubric rewards over runs that finished.

Model Ladybird Chairs AirNow
GPT 6 Astra 93.3% 0% 100%
Claude Fable 5.1 56.0% 0% 100%
Claude Sonnet 5 0% 0% 100%
Qwen 3.8 Max* 0% 16.7% 95.0%

* Qwen ran with shell and file tools but no computer-use tool. Its one Chairs success came from a run that installed YOLO object detectors to count the chairs.

Run them

uv tool install hud --python 3.12 --force
hud set HUD_API_KEY=your-key   # from hud.ai settings
hud eval <taskset-name> claude --runtime hud --all --max-steps 300 -y

To copy these tasks into your own taskset:

hud sync tasks my-liveweb-samples tasks.py

.hud_eval.toml re-encodes screenshots as WebP, because full-size screenshots of the 3D viewer overrun the provider's request-size cap on long runs.

License

MIT

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages