Can this model do qa tasks？

[leexinhao]

I find that the model seems to have a hard time output options, my prompt is：

Detect and report the timestamp of the video frame that semantically matches the given textual query "puts the bread on the plate ". 
Choose the correct option to the following question: 
(0) 33.59402815508409, (1) 37.721428155084084, (2) 39.995228155084085, (3) 50.98892815508409, (4) 55.96002815508409, 
Best Option: (

[RenShuhuai-Andy]

Hi, thanks for your interest.

I think you should do some prompt engineering, like:

text_input = "In the context of a culinary event where someone 'puts the bread on the plate', determine the closest time from the following options:"\
"(A) 33.5 seconds; (B) 37.7 seconds; (C) 39.9 seconds; (D) 50.9 seconds; (E) 55.9 seconds."\
"Please select the correct option by responding with the corresponding letter (A)-(E). Avoid outputting anything other than the designated letters."

The output is:

The closest time is (B) 37.7 seconds.

[RenShuhuai-Andy]

@leexinhao Hi, we found that reduce lora_alpha from 32 to 20 during inference can be very helpful for qa tasks. See https://github.com/RenShuhuai-Andy/TimeChat/blob/master/docs/FAQ.md#3-how-to-better-instruct-the-model-to-perform-qa-or-other-specialized-tasks