README.md

metadata

task_categories:
  - visual-question-answering
language:
  - en
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: data
        path: data/data-*
dataset_info:
  features:
    - name: question_ref
      dtype: string
    - name: images
      list: string
    - name: question_text
      dtype: string
    - name: expected_answer
      dtype: string
    - name: map_count
      dtype: string
    - name: spatial_relationship
      dtype: string
    - name: answer_type
      dtype: string
    - name: domain
      dtype: string
    - name: map_elements
      list: string
    - name: context_images
      list: string
  splits:
    - name: data
      num_bytes: 576010
      num_examples: 500
  download_size: 120231
  dataset_size: 576010
pretty_name: FRIEDA

FRIEDA is a multimodal benchmark for open-ended cartographic reasoning over real-world map images.
Each example pairs reference maps (and optional contextual maps) with a natural-language question and a reference answer. The benchmark targets common GIS relation types (i.e., topological, metric, directional) and includes questions that require multi-step reasoning and cross-map grounding.

Dataset Summary

• Modality: image + text
• # Examples: 500
• Input: map image(s) + question text
• Output: expected answer (textual)
• Metadata: map_count, domain, relationship type, map elements

Languages

The dataset questions and answers are in English.

---

How to use it

from datasets import load_dataset

# Full dataset (split name = "data")
ds = load_dataset("knowledge-computing/FRIEDA", split="data")
print(ds[0].keys())
print(ds[0]["question_text"])    # Actual question being asked
print(ds[0]["images"])           # List of string paths to images (e.g., "images/...png")
print(ds[0]["context_images"])   # List of string paths to contextual images