example_hists/README.md

Example Histograms with VQA

Directories

• example_hists/imgs stores the example images (right now there are PDF and JPEG formats)
• example_hists/jsons stores the jsons

The Data

Basics

To load json (after setting json_dir to where the jsons are stored):

import json
data_file = json_dir + 'nclust_4_trial8.json'
with open(data_file,'r') as f:
    t = json.load(f)
    datas = json.loads(t)

VQA can be accessed with the VQA tag. For example, datas['VQA'] prints out the same Q and A pairs as before:

{'Level 1': {'Figure-level questions': {},
  'Plot-level questions': {'nbars ': {'plot0': {
    'Q': 'How many bars are there on the figure? You are a helpful assistant, please format the output as a json as {"nbars":""} for this figure panel, where the "nbars" value should be an integer.',
     'A': {'nbars ': 50}}}}},
 'Level 2': {'Plot-level questions': {'minimum (plot numbers)': {'plot0': {
    'Q': 'What are the minimum data values in this figure? You are a helpful assistant, please format the output as a json as {"minimum x":""} where the minimum value of "x" is calculated from  the data values used to create the plot in the format of floats.  ',
     'A': {'minimum (plot numbers)': {'minimum x': 0.37513605039268844}}}},
   'minimum (words)': {'plot0': {'Q': 'What are the minimum data values in this figure? You are a helpful assistant, please format the output as a json as {"minimum x":""} where the minimum value of "x" is calculated from  the data values used to create the plot in the format of floats.  ',
    ...
        }
      }
    }
  }
}

Additional question breakdowns

Additionally, there are several extra tags for a breakdown of the question following Microsoft's Elements of a Good Prompt slide which are at the same "level" as the A and Q of the json example above.

• persona -- who the AI is pretending to be, along the lines of "you are a helpful assistant...", this is often the "system" prompt when passing to APIs like ChatGPT/Claude
• context -- what does the tool need to know, for single plots like these single example histograms this is empty, but for multi-panel plots, the layout of the figure will be specified and the specific plot pointed to
• question (Microsoft's "objective") -- the actual question (e.g., "what is the mean of the distribution?")
• format (Microsoft's "boundaries") -- specify the output format, something like "export as a json with {"nbars":""}..."

Question "Levels"

Here "levels" refer to the 4-level paradigm developed in Accessible Visualization via Natural Language Descriptions: A Four-Level Model of Semantic Content.

Like the prior iteration, there are different "levels" of parsing the plots, as well as figure-level and plot-level questions, assuming each figure object can be made up of multiple plot axes objects (right now, there is a single axes, so just "plot0" for everything).

For example, to access the plot-level, Level 1 questions:

datas['VQA']['Level 1']['Plot-level questions']

prints out:

{'nbars ': {'plot0': {'Q': 'You are a helpful assistant that can analyze images.  How many bars are there in the specified figure panel? Please format the output as a json as {"nbars":""} for this figure panel, where the "nbars" value should be an integer.',
   'A': {'nbars ': 50},
   'persona': 'You are a helpful assistant that can analyze images.',
   'context': '',
   'question': 'How many bars are there in the specified figure panel?',
   'format': 'Please format the output as a json as {"nbars":""} for this figure panel, where the "nbars" value should be an integer.'}}}

Example LLM Outputs

Various example outputs from LLMs will be in the LLM_output subfolder in this directory.