Zendo Synthetic Visual Reasoning Dataset

Synthetic Zendo-style scenes with associated rules and per-scene tensor representations. Each scene either follows ("positive", label=1) or violates ("negative", label=0) a rule that is given in natural language and as a Prolog query.

Splits

Layout

• images/<split>/<batch>/<rule_id>/<scene_id>.png — rendered scene
• tensors/<split>/<batch>/<rule_id>/<scene_id>.pt — pre-computed scene tensor
• metadata/scenes_train.parquet, metadata/scenes_test.parquet — main tables (one row per scene, with image_path, tensor_path, label, rule_nl, rule_query)
• metadata/rules.parquet — one row per (batch, rule_id) with rule_nl, query_pos, query_neg

Loading

from datasets import load_dataset
ds = load_dataset("ss567uhg/zendo-synthetic-data")
ds["train"][0]

Tensors are not auto-loaded by datasets; load them via:

import torch
scene = ds["train"][0]
t = torch.load(scene["tensor_path"]) if scene["tensor_path"] else None

Citation

TBD.