Can't to see the performance of GPT4-V on DESCRIPTION and ASSESSMENT tasks! about q-bench HOT 1 CLOSED

These will be put into a further plan.

1. For the DESCRIPTION task, one concern is that the current pipeline using GPT-3.5 to evaluate GPT-4V may not be enough accurate, so we are working on human_eval to evaluate it (same for open-source models). Please stay tuned as that may take time.

2. For the ASSESSMENT task, one big problem is that GPT-4V does not support logits as outputs, which is the basis of our evaluation setting. Another concern is that given the IQA datasets are large, without API, developing a good strategy for it and further testing will be very difficult and time-consuming. We will take a deep dive into this when the GPT-4V API is out.

from q-bench.