We have to have a suite of checks to assess how good the LLM captions are. Here is an example scene:
And here is one of its captions from Qwen:
"Jump carefully between isolated islands while avoiding enemies that guard the final stretch."
We can objectively verify that there are no enemies in this scene. This is one way of confirming that the caption does not describe the scene well. Some digging is required to see if there might be issues with the prompts that are leading to this bad data generation, but eventually we have to assess the captions the LLMs give us, and this caption is just incorrect in claiming that something is present when it is not.
Similarly, the caption neglects to mention objects that are present, such as question blocks and the mushroom.
Basically, there are two parts to this issue:
- There might be ways we can do better prompt engineering to encourage the models to caption this correctly. Maybe we should tell the models that "all object types that are present must be mentioned"
- However, regardless of whether we can improve the captions with better prompting, we need deterministic, objective ways of assessing the quality of the LLM-generated captions. For each type of thing that can be present in a level, we can check if the caption correctly or incorrectly claims that such objects are present. If the caption makes a claim about precise counts, then we can check that claim. This will be tricky, and perhaps a bit error prone, but can still be done deterministically. The only thing I worry about are synonyms for objects. Perhaps in developing a prototype of this, we can start by having the code scan all captions and identify nouns (or maybe more broadly, words) that don't correspond to any tile/object that it is aware of. Perhaps you can make a GUI that scans the data and shows the scenes, and mentions words in the caption that it thinks are not represented, and if the human notices that a synonym was used, we can provide that information (essentially making a thesaurus to use for calculations).
We have to have a suite of checks to assess how good the LLM captions are. Here is an example scene:
And here is one of its captions from Qwen:
"Jump carefully between isolated islands while avoiding enemies that guard the final stretch."
We can objectively verify that there are no enemies in this scene. This is one way of confirming that the caption does not describe the scene well. Some digging is required to see if there might be issues with the prompts that are leading to this bad data generation, but eventually we have to assess the captions the LLMs give us, and this caption is just incorrect in claiming that something is present when it is not.
Similarly, the caption neglects to mention objects that are present, such as question blocks and the mushroom.
Basically, there are two parts to this issue: