Question about get_target_obs_and_act

[jim850223]

Hi,
I am curious about in get_target_obs_and_act of env.py. My question is: why would you check whether the ground pos_candidates is not in the cleaned_html? Since the mechanism is adopted when the agent is being evaluated, doesn't it mean that the agent check the ground truth to modify the way it interact with the bench mark?

[zorazrw]

The env.py is directly adopted from the Mind2Web project. In the current implementation, this function is only used when constructing few-shot examples but not during agent prediction.