browsergym-assistantbench 0.14.3.dev1

AssistantBench benchmark for BrowserGym

AssistantBench <> BrowserGym

This package provides an implementation for using the AssistantBench benchmark in BrowserGym.

Because AssistantBench includes open-ended tasks, setup is extremely easy and simply requires installing the package.

Please note that AssistantBench has a hidden test set, so test set predictions will need to be uploaded to the official leaderboard.

Setting up

• Install the package (this is still a wip)

pip install browsergym-assistantbench

• Run inference, e.g., run the following commands for demo on a simple toy task

python demo_agent/run_demo.py --task_name assistantbench.validation.3

• Test set predictions will be saved to ./assistantbench-predictions-test.jsonl. To evaluate on the official test set, upload these predictions to the official leaderboard.