smokedataset_QA / README.md
FranciscoLozDataScience's picture
forgot comma
aa11ab1
metadata
license: mit
task_categories:
  - visual-question-answering
task_ids:
  - multi-label-image-classification
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': cloud
            '1': other
            '2': smoke
    - name: prompt
      dtype: string
    - name: choices
      sequence: string
  splits:
    - name: test
      num_bytes: 119949703
      num_examples: 19832
  download_size: 132474880
  dataset_size: 119949703
tags:
  - climate

Motivation

My goal is to build a dataset using Wild Sage Node captured images to help score LLMs that will be used with SAGE.

Origin

This dataset was forked from sagecontinuum/smokedataset

Data Instances

A data point comprises an image, its classification label, a prompt, and mulitple choices.

{
  'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224 at 0x1215D0C50>,
  'label': 2,
  'prompt': 'What is shown in the image?',
  'choice': ['cloud', 'other', 'smoke']
}

Data Fields

  • image: A PIL.JpegImagePlugin.JpegImageFile object containing the image.
  • label: the expected class label of the image.
  • prompt: the prompt that will be sent to the LLM.
  • choice: the choices that the LLM can choose from.

Scoring

The multiple choice portion of the question is scored by overall accuracy (# of correctly answered questions/total questions). The question can also be open-ended by eliminating the choice portion.

Next Steps

More work is needed to figure out a scoring for open ended questions.

Citation

Dewangan A, Pande Y, Braun H-W, Vernon F, Perez I, Altintas I, Cottrell GW, Nguyen MH. FIgLib & SmokeyNet: Dataset and Deep Learning Model for Real-Time Wildland Fire Smoke Detection. Remote Sensing. 2022; 14(4):1007. https://doi.org/10.3390/rs14041007