Why doesn’t VizWiz’s `build_prompt` include the short-answer instruction like `ImageVQADataset`?

I noticed that in ImageVQADataset you override build_prompt to append:

Answer the question using a single word or phrase.

This ensures the model knows to produce a short answer. However, the VizWiz class inherits build_prompt directly from ImageBaseDataset and doesn’t include this instruction, which leads to long descriptive outputs that always score 0 under VQA Accuracy.

Could you clarify why this short-answer prompt is added for ImageVQADataset but not for VizWiz? Is there a reason for treating VizWiz differently, or would you consider adding the same instruction to improve scoring?

Empirical Observation:
After manually adding this line to VizWiz.build_prompt, Ovis2-8B’s VQA Accuracy jumped from 3.09% to 39.77% on the VizWiz evaluation set.

Thank you!

Compare ImageVQADataset.build_prompt with the inherited ImageBaseDataset implementation used by VizWiz. Check the VizWiz evaluation set and VQA Accuracy behavior after confirming the intended prompt format. Done means the project has a documented or implemented decision for whether VizWiz should receive the short-answer instruction, with evaluation results supporting it.