metadata
license: mit
task_categories:
- visual-question-answering
- image-to-text
language:
- en
tags:
- vision-language-model
- benchmark
- causal-reasoning
- scientific-visualization
- multimodal
pretty_name: VeriRender
size_categories:
- n<1K
dataset_info:
features:
- name: sample_id
dtype: string
- name: split
dtype: string
- name: domain
dtype: string
- name: family
dtype: string
- name: clean_image
dtype: image
- name: corrupted_image
dtype: image
- name: eval_image
dtype: image
- name: perturbation_type
dtype: string
- name: perturbation_description
dtype: string
- name: prompt
dtype: string
- name: generator_content
dtype: string
- name: generator_type
dtype: string
- name: seed
dtype: int64
configs:
- config_name: default
data_files:
- split: inconsistent
path: manifest_inconsistent.parquet
- split: consistent
path: manifest_consistent.parquet
VeriRender Benchmark Dataset
Causal consistency verification samples for Vision-Language Models.
Layout
manifest.jsonl ← canonical index (one row per sample)
benchmark.yaml ← config used to generate this release
inconsistent/{domain}/{sample_id}/ ← corrupted evaluation samples
consistent/{domain}/{sample_id}/ ← negative controls (clean images)
Splits
| Split | Description | Eval image |
|---|---|---|
inconsistent |
Symbolic spec is correct; image has a perturbation | corrupted.png |
consistent |
Symbolic spec matches the clean image | clean.png |
Sample folder
Each sample contains:
spec.py/spec.tex/spec.txt— symbolic generator (unchanged for inconsistent samples)clean.png— faithful renderingcorrupted.png— perturbed rendering (inconsistent only)prompt.md— VLM evaluation promptmetadata.json— full provenance
Loading
import json
from pathlib import Path
root = Path(".")
rows = [json.loads(line) for line in (root / "manifest.jsonl").open()]
Or rebuild the manifest after edits:
python scripts/build_manifest.py