IamBusy/OpenJev-Vision-Research-v0.1

OpenJev Vision Research v0.1

12,832 image records, with public provenance, original synthetic scenes, and programmatically derived decision questions.

This is an experimental research dataset for visual posterior learning and compositional decisions, released with OpenJev. It is not a reproduction of TypeSafe's proprietary Jev model or training method.

Three separate configurations

Config	Images	What the labels mean	License
synthetic	8,192	Exact posterior over 64 worlds under a specified noisy sensor and supplied prior	Apache-2.0
pets	2,960	Original breed/species annotations of real photographs	CC BY-SA 4.0
clevr4	1,680	Original attributes of public rendered scenes; 20 color-shape combinations withheld from training	CC BY 4.0

Do not treat class labels in pets/clevr4 as ground-truth probability distributions. Do not treat the synthetic posterior as a real-world calibration guarantee. No closed-model teacher was used to assign any released label.

Loading

from datasets import load_dataset
import json

data = load_dataset("IamBusy/OpenJev-Vision-Research-v0.1", "synthetic")
example = data["train"][0]
image = example["image"]
questions = json.loads(example["questions_json"])
targets = json.loads(example["target_json"])

Choose pets or clevr4 for the public-image configurations.

Schema

Each record contains a decoded image, unique id, group_id, source, source_id, split, license, and three JSON strings:

• questions_json: typed questions, executable event semantics and derived targets.
• target_json: synthetic prior/posterior and latent-world annotation, or public class labels.
• provenance_json: generation parameters or original source hashes and transformation records.

Synthetic observed metadata and latent truth are supervision/oracle fields; the vision models must not read them at inference. Model inputs are image pixels and the explicitly supplied prior. Event execution uses a fixed, documented ontology. Descriptive instructions are not evidence of learned free-form language understanding.

Synthetic construction

Three fixed slots (left, center, right), each with blue/red color and circle/square shape, define 64 worlds. A declared six-variable Ising prior is supplied. Each attribute is flipped independently with probability 0.08 by a synthetic sensor. Images render observed attributes; whole objects may be occluded. The exact posterior is computed from observable sensor values, not set to the hidden world's one-hot label.

Each episode has one occluded and one fully visible view of the same sensor observation. Their split and appearance seed remain shared. The 36 Boolean questions per image yield 294,912 derived targets, not independent samples. Pixel texture is independent appearance noise. Training priors use local/chain edges; test_topology changes dependency structure; test_appearance changes the rendering palette. Finite world assignments can repeat across splits.

Public source selection

Oxford-IIIT Pet: original data by Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman and C. V. Jawahar, Cats and Dogs, CVPR 2012. Official dataset and license. Downloaded through timm/oxford-iiit-pet at revision 089695c834a7deb60505b7cc506672db1c31a6aa; original IDs, breed names, species and splits were checked against the original Oxford annotation archive. Per breed: 40 train, 10 dev, 10 calibration from official trainval; 20 official test images for test. Original image owners retain copyright.

CLEVR-4: original data by Sagar Vaze, Andrea Vedaldi and Andrew Zisserman, No Representation Rules Them All in Category Discovery, NeurIPS 2023. Official dataset and license. Selected from the official 10k v1 archive. Colors and shapes are sorted lexically. Pairs for which (color_index + shape_index) modulo 5 equals zero are absent from train/dev/calibration. For the other 80 pairs, select 12/2/2 original train images. Test has four original validation images for each of all 100 pairs. This is not the full official benchmark.

Selection uses a fixed SHA256 ordering of original IDs. Public images are converted to RGB, resized so the longest side is at most 256 pixels using Lanczos, and saved as lossless PNG. Original IDs and hashes remain in every row. Questions are newly derived from existing annotations using documented templates; no new visual attributes were hallucinated or inferred from a language model.

Splits and leakage boundaries

Config	Train	Dev	Calibration	Test
synthetic	4,096	512	512	1,024 each: ID, topology, appearance
pets	1,480	370	370	740
clevr4	960	160	160	400

All derived views/questions for an original source image or synthetic episode stay in one split. Exact selected-image duplicates across splits were checked. No claim is made about eliminating all near duplicates or foundation-model pretraining contamination. Synthetic and public results must be reported separately.

Reproducibility

The three source manifests and EXPORT_MANIFEST.json provide source revisions, selection rules and SHA256 checksums for the Parquet files. Code and protocols are in OpenJev's visual research track. Evaluate compound probabilities and held-out compositions, not just format validity or in-distribution top-1 accuracy.

Limitations

Fixed small ontologies, synthetic sensor assumptions, and bounded public subsets limit conclusions. CLEVR-4 objects within an image share color and shape, so it does not establish arbitrary object-binding ability. Photo labels describe the annotated pet, not every object in a scene. Some probabilities can be confidently wrong. This dataset is not a safety certification or evidence of frontier capability.

Licensing

Licenses are source-specific. Oxford-derived images and their adaptations remain CC BY-SA 4.0. CLEVR-4-derived images remain CC BY 4.0. Original synthetic scenes and original generator code are Apache-2.0. See LICENSES.md; do not relabel the whole mixture Apache-2.0.