This repository is the implementation of AutoGuide (https://arxiv.org/pdf/2403.08978).
The webarena environment is built on top of WebArena. Follow WebArena to set up the environment, or reproduce via:
conda env create -f environment.yaml
conda activate autoguide
playwright installHuman demonstrations are stored as pickle files under extracted_data/reddit/<task_id>/:
human.pkl: Demonstrations with action sequences (no reasoning, used in original experiments).human_reason.pkl: Demonstrations with action sequences and reasoning.
Both formats are supported.
To run both guideline extraction and evaluation at the same time:
python main.py --model gpt-4-1106-preview --result_dir ./exp_logs --apply_guidelines --extract_guidelinesUsing --apply_guidelines together with --extract_guidelines runs both extraction and evaluation simultaneously.