The code to upload MMDT's data onto huggingface
Contain 6 perspectives: adversarial robustness, fairness, privacy, hallucination, ood and safety.
origin_attribute: the original attribute of data in split "attribute"
origin_object: the original object of data in split "object"
origin_relation: the original relation of data in split "spatial"
object_a, object_b: the objects used in split "spatial", template is {object_a} {relation} {object_b}
object: the object used in split "attribute", template is {attribute} {object}
label: the label of the data
surrogate_model: the model attacked by the algorithm
algorithm: the algorithm used to generate this data
q_gender: questions about gender
q_race: questions about race
q_age: questions about age
task: street_view or selfies
type_street_view: the specific difficulty of the street_view task, single/group & text/no text
country, state_province, city, latitude, longitude, zipcode: the label of data in split "street_view"
ethnicity: caucasians or hispanics, only works in split "selfies"
label_selfies: the label of data in split "selfies"
type_selfies: ID or Selfie, distinguish the type of image in split "selfies"
question: prompt to query
id:
task:
label: in cooccurence part
target: in cooccurence part
keyword: in misleading and ocr part
answer, bbox, natural_question, natural_answe: in counterfactual, distraction and natural part
id: data id
img_id: image id
question:
answer:
task: attribute, count, identification and spatial