Clarification about prompts and Result in Table 4

[kushagra414]

Hi,
I have some doubts about the paper, it would be grateful if you could clear them up:

1. Could you please provide the system prompt and the user prompt you used as the input to Llama-2-7B?
2. What parameters did you set in llama-2-7b, temperature, top_p, etc. It would be helpful if you could provide the code.
3. Did you the Llama-2-7B-chat or Llama-2-7B to generate descriptions whose results are mentioned in Table 4, in the paper?

[Reza-esfandiarpoor]

Hi,

To generate the results of the original Llama model, we use the following prompt:

You are a knowledgeable teacher.

For the following objects, generate captions that represent the distinguishing visual differences between the photos of the two objects.
Generate as many captions as you can.

Object 1: Field sparrow
Object 2: acadian flycatcher

Visual characteristic: Bill shape
Caption 1: A photograph of a field sparrow, a type of bird, with cone bill shape.
Caption 2: A photograph of a acadian flycatcher, a type of bird, with all-purpose bill shape.

Visual characteristic: Breast color
Caption 1: A photograph of a field sparrow, a type of bird, with buff breast color.
Caption 2: A photograph of a acadian flycatcher, a type of bird, with white breast color.

Visual characteristic: Back pattern
Caption 1: A photo of a field sparrow, a type of bird, with striped back pattern.
Caption 2: A photo of a acadian flycatcher, a type of bird, with solid back pattern.

For the following objects, generate captions that represent the distinguishing visual differences between the photos of the two objects.
Generate as many captions as you can.

Object 1: CLASS_NAME_ONE
Object 2: CLASS_NAME_TWO

We use these parameters:

top_k = 10
max_length = 1024
temperature = 0.1
top_p = 0.1
do_sample = True

We use Llama 2 - 7B in our experiments.