---
dataset_info:
  features:
    - name: image_id
      dtype: string
    - name: image
      dtype: image
    - name: instance_idx
      dtype: int64
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: instance_triplet
      dtype: string
    - name: triplets
      dtype: string
    - name: all_object
      dtype: string
    - name: object
      dtype: string
  splits:
    - name: train
      num_examples: 1226
---

# Tri-HE

A triplet-based hallucination evaluation dataset for visual question answering. Contains 1226 flattened question-answer instances across 300 images. Each row includes the question, answer, and associated knowledge triplets (both instance-level and image-level) encoded as JSON strings. The triplets capture subject-predicate-object relationships used to probe visual hallucination.