XInstructBLIP demo text generation

[ParkJun-Yeong]

hi. i executed demo.ipynb setting modality to only 'audio'.

in vicuna7b_v2.yaml

  modalities: ['audio']

and I run the inference function in demo.ipynb because I don't need the web UI.

inference(None, None, "examples/audio/Group_of_Dogs_Barking.wav", None, "Question: What is the dog doing? Answer: ", "Qformer Prompt: What is the dog doing?", 1, 250, 5, 1, 1.5, 0.9, "Beam search")

inference(None, None, "examples/audio/Group_of_Dogs_Barking.wav", None, "Question: What is the dog doing? Answer: ", "Question: What is the dog doing? Answer: , 1, 250, 5, 1, 1.5, 0.9, "Beam search")

all config for generation was set samely as original demo code.
but in both 'beam search' and 'nucleus sampling', the generated output was below.

how can I reproduce the result of the paper?