WebCoach

Setup: browser-use and docker

First cd ~ and then clone the repository (or fetch it if you've already cloned it):

git clone https://github.com/browser-use/browser-use.git
cd browser-use

git fetch --tags

git checkout tags/0.7.1 -b my-0.7.1-branch

Then install docker image on this machine:

First change one line in ~/browser-use/docker/build-base-images.sh

$build_cmd $tag_args $build_args -f $dockerfile ..

Then follow ~/browser-use/docker/README.md for docker image installation

Setup: running local models

Follow the scripts under ~/start_LLM_inference. SGLang provides their own docker image already.

Running the Benchmark

First make sure you configurate the experiments here using ~/run_benchmark/config.yaml

Then follow ~/run_benchmark/run_webvoyager.sh

You can specify base_dir, model, benchmark, and coaching parameters

Evaluation

Run ~/evaluation/eval_webvoyager_results.py to generate table, metrics include SR, Succ/Total, Avg_time, Avg_num_steps.