benchmarkanon/logic_conflict

LogicCon

LogicCon contains real photographs paired with statements that are consistent with or conflict with a grounded visual fact. It supports image–statement conflict detection, conflict classification, and inspection of the underlying evidence. This is an anonymous release of version 2.0.0.

Statements were generated with controlled templates and screened with Qwen2.5-VL-7B-Instruct. Images are original GQA photographs, not generated images. The synthetic tag refers to statement construction.

Contents and loading

from datasets import load_dataset

ds = load_dataset("benchmarkanon/logic_conflict")
example = ds["test"][0]
image, statement = example["image"], example["statement"]

The default dataset exposes only sample_id, image, and statement. Gold labels and evidence are in annotations/{train,dev,test}.jsonl, joined by sample_id. Do not pass the gold annotations, source questions, paired statements, or screening explanations to a model during evaluation.

LogicCon-anonymous.zip is the complete offline bundle: images, input JSONL, gold annotations, selected source questions/scene graphs, reproduction code, and checksums. Extract it to obtain a LogicCon/ directory. See USAGE.md for loading and verification instructions. The ZIP and Parquet represent the same examples and byte-identical images; downloading both is optional.

Construction and splits

The source is GQA's official balanced train/validation questions (v1.2), functional programs, and scene graphs (v1.1). The pipeline executes supported program operations, resolves object references, renders a supported statement, and constructs a controlled contradictory statement. Attribute, category, relational, and spatial mutations have separate constraints.

Candidates underwent VLM screening and limited direct AI inspection. Deterministic selection retained eligible pairs under split, duplicate, and per-image constraints.

Source GQA validation images form the test split. Source training images are assigned to train/dev by image-content hashing and selected to achieve the released split counts. No identical image files or identical decoded RGB images cross these splits. This does not exclude perceptual near-duplicates.

provenance/replay_candidates.jsonl.gz contains the selected source question/scene-graph records. provenance/image_manifest.jsonl maps image hashes to GQA image IDs. provenance/source_manifest.json records official annotation archive URLs and hashes. provenance/screening_config.json records model/prompt hashes, seed, decoding settings, and a public model ID. Historical intermediate fields in replay records describe their original construction stage; final generation is the within-image, constrained-template procedure stated here.

Limitations

• Single-model visual screening is fallible and can favor the screening model. Its screening decisions are not an independent performance estimate.
• Negation and template cues remain, and text-only diagnostics found exploitable shortcuts. Report a text-only baseline when evaluating.
• Candidate construction uses deterministic source order and caps, rather than uniform sampling of all GQA questions.
• Source annotation errors, visual ambiguity, perceptual near-duplicates, and pretraining contamination may remain.
• Public gold labels and paired examples are intended for reproducible research; this release is not a hidden-label evaluation service.

Attribution and terms

GQA: Drew A. Hudson and Christopher D. Manning, GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering, CVPR 2019. Paper, official source.

Visual Genome: Ranjay Krishna and collaborators, Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations. Paper, official project mirror.

Screening model: Qwen/Qwen2.5-VL-7B-Instruct. Model weights are not distributed here.

See LICENSE.md for the distinction between new annotations/code and upstream material. Authorship of this benchmark is withheld during anonymous review; upstream credit is retained. ANONYMITY.md describes the scope of the publication checks. This release contains no paper PDF or claimed large-model evaluation results.