Train/val split manifests extracted from GeoChat_Instruct
for fine-tuning vision-language models on satellite scene classification.
45 scene categories (e.g. airplane, baseball diamond, beach, bridge, …)
Split
Samples
train
~25 200
val
~4 725
Each JSON file is a list of objects with fields:
image (relative path), question, answer (class label).
