Prompt for Screenspot and v2

[SunzeY]

Thank you for this awesome work! however, I cannot reproduce results of Qwen2.5-VL 3b/7b baseline results on ScreenSpot and ScreenSpot-v2 as reported in the paper. I have used the prompt at here. Can you share the prompt used for evaluation on these benchmarks? Thank you!

[MillanK0817]

Hi,

Thank you for your feedback. Unfortunately, we didn't retain our evaluation script for Qwen2.5-VL. However, we recently evaluated Qwen2-VL performance using the following prompt:

prompt = f"In this screenshot, tell me where I should click to complete the following task: \"{contexts}\"\n\n" \
            f"You should directly provide your answer in the format [x1, y1, x2, y2] using relative coordinates between 0 and 1000. Do not include any explanation or additional content."

We used a straightforward approach without complex prompting for Qwen2.5-VL, employing similar prompts to the one shown above.

There's a key difference in coordinate systems between the models:

Qwen2-VL uses relative coordinates, while Qwen2.5-VL uses absolute coordinates

You'll need to adjust your prompts accordingly to account for this coordinate system difference.