knowledge-computing/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