use old prompt version

2. Avoid walls (#).

Actions you can take: Up, Down, Left, Right. You can take up to {max_action_per_step} action(s) at a time.

- Up: move up to the cell above

- Down: move down to the cell below

- Left: move left to the cell to the left

- Right: move right to the cell to the right

If there is a box on the cell you want to move to, you will push the box one cell in the same direction.

Up: move up to the cell above (to the above row).

Down: move down to the cell below (to the below row).

Left: move left to the cell to the left (to the left column).

Right: move right to the cell to the right (to the right column).

Rewards:

Box on target: +1.0

Format correct: +0.5

Please think step by step and provide the actions you want to take.

You should wrap your thought between `<think>` and `</think>` tags, and wrap your answer between `<answer>` and `</answer>` tags.

Your response should STRICTLY follow the format:

<think>...</think><answer>...</answer>

Your reponse should be in the format of <think>...</think><answer>...</answer>

"""

# E.g. <think> There's a box on the upper right of me, the target is on the upper side of the box, I need to push the box it upward. </think><answer> Right,Up,Up </answer>

# Let's try to use a format reward and answer reward

[Initial Observation]:

{observation}

Decide your next action(s).

Your response should STRICTLY follow the format:

<think>...</think><answer>...</answer>

Your reponse should be in the format of <think>...</think><answer>...</answer>

"""

action_template = """Valid action extracted from your response is {valid_action}.\

action_template = """After your answer, the extracted valid action is {valid_action}.\

After that, the observation is:

{observation}

reward: {reward}

done: {done}

Decide your next action(s).

Your response should STRICTLY follow the format:

<think>...</think><answer>...</answer>

Your reponse should be in the format of <think>...</think><answer>...</answer>

"""

 No newline at end of file

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=10 \

    trainer.test_freq=20 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=5 \