Input
image raw path
question raw
options raw
question_with_extra raw
extra_info (provenance / answer_type / task) derived
format note LLaVA conversation
Output
ground-truth answer raw
expected response format conventional
Varies by task: option letter (mcq), yes/no (judgement), scalar+unit (depth/size), or structured coords (3D detection).
evaluation from paper
Training target (SFT) — the gpt turn value, used directly as the supervision label.
Raw record from SpatialCorpus jsons/ · LLaVA entry (mapped)