shbiao/ImageText-Conflict-Screening-Evaluation

ImageText Conflict Screening Evaluation

Resources for From Pairwise Signals to Screening Decisions: Supervised Readouts and Proxy Misalignment in Image–Text Conflict Detection.

Code and reproduction · Data dictionary · Component licenses

Dataset Summary

This resource connects image–text compatibility judgments with frozen VLM decision scores. It provides 1,050 adjudicated conditions from 525 images, anonymized independent reviewer labels, model-contact exclusions, supervised-readout predictions, controlled architecture–supervision results, and all 7,511 SugarCrepe caption comparisons. Cached hidden representations support CPU readout fitting and checkpoint verification.

Annotation and Evaluation Data

Resource	Size	Role
Random human-reference cohort	350 images / 700 conditions	Reassessment under visual judgments
Diagnostic human-reference cohort	175 images / 350 conditions	Targeted difficult conditions
Common training/development exclusion	245 images / 490 conditions	Shared six-model comparison scope
Architecture–supervision matrix	80 fits	Four readouts × four scales × five seeds
SugarCrepe	7,511 pairs / 1,560 COCO images	Native compositional caption selection

The two human-reference cohorts retain their sampling identities. CLASH source-original/source-edited labels and adjudicated visual judgments are separate fields.

Dataset Structure

data/           Human judgments, source labels, exact captions, split/contact manifests
predictions/    Frozen original/readout/external scores
statistics/     Point estimates, paired intervals, error layers, calibration summaries
features/       Cached h vectors and row-to-caption mappings
protocols/      Model versions, scoring, training schedules, field and license definitions
checksums/      Source provenance and SHA-256 release manifest

case_id joins human tables. Historical model tables join on image ID and condition within their named cohort. pair_id joins SugarCrepe tables. Captions remain exact; no COCO image file is redistributed. data/case_metadata.csv gives COCO source references, caption SHA-256, encoded-image SHA-256, and decoded-pixel hashes for the human-reference inputs.

Human-reference Annotation

Two reviewers independently assigned CONFLICT, COMPATIBLE, or INSUFFICIENT from the image and exact caption. They agreed on 941/1,050 conditions (89.62%; Cohen's κ=0.7984). All 109 disagreements received final adjudicated labels. The retained records do not establish a separate independent third adjudicator. The frozen reference contains 495 conflicts, 551 compatible conditions, and four insufficient conditions. Reviewer columns retain labels under generic identifiers, with personal identities and internal annotation records excluded.

The primary shared reassessment excludes the union of model training and development images: 245 images remain, including 162 random and 83 diagnostic images. The random subset has 321 effective binary conditions after three insufficient judgments. split_manifests.csv also supports model-specific exclusion. These flags capture model contact; the full annotation collection remains available for descriptive analyses.

Model Predictions

Six models are retained on the human reference: original Classification QLoRA score, Linear-1024, Residual-1024, expanded Residual-1536, adapted LoRA with the old head, and adapted LoRA with the joint head. The model registry identifies adapter and readout checksums, standardizers, full-label scoring, and frozen thresholds. The released predictions and cached features cover the primary Qwen experiments. Frozen Qwen classification and joint-adaptation LoRA checkpoints, supervised readout weights, and standardizers are publicly distributed through GitHub Releases v0.1.0, as model-assets-v1.zip. The Dataset contains no model checkpoints; neither resource redistributes Qwen base-model weights or original COCO images.

On the original readout evaluation cohort, mean pAUC increases from 0.6693 to 0.7671 for Linear and 0.7585 for Residual. On SugarCrepe's native protocol, Linear improves caption choice from 95.06% to 96.14%. Complete scores retain alternatives and negative results alongside these findings.

Evaluation Protocol

Higher score means greater conflict. Binary decisions use the model's historical development threshold with strict score > threshold. AUROC and integral-normalized pAUC summarize ranking; ROC recall at FAR≤5% differs from recall after transferring a frozen development threshold. Whole-image paired bootstrap preserves sibling conditions.

Human INSUFFICIENT judgments remain recorded and are excluded from binary metrics; source/human comparisons use the same effective conditions. Random and diagnostic cohorts are reported separately. SugarCrepe selects the caption with lower conflict score; ties and failures count as incorrect. The CLASH-derived binary task and SugarCrepe caption-choice task have separate evaluation identities from CLASH's official question-driven QA protocol.

Reproducing Paper Results

hf download shbiao/ImageText-Conflict-Screening-Evaluation --repo-type dataset --local-dir resources
git clone https://github.com/SHbiao/ImageText-Conflict-Screening.git
cd ImageText-Conflict-Screening
python -m pip install -r requirements-evaluation.txt
python scripts/reproduce.py --data ../resources --output outputs/paper-tables

The Dataset is public and ungated; downloading does not require a login.

This CPU entry point reconstructs the main tables directly from predictions, including all 72 human/source metric rows and all 32 SugarCrepe category/model rows. checksums/sha256.json verifies the local resource version.

Related Resources

• CLASH: source captions and edit metadata.
• COCO: original image sources; retrieve images under their own terms.
• SugarCrepe: external compositional caption pairs.
• Qwen3-VL-4B-Instruct: base VLM.

Citation and License

From Pairwise Signals to Screening Decisions: Supervised Readouts and Proxy Misalignment in Image–Text Conflict Detection. Unpublished manuscript prepared for Multimedia Systems. Software citation metadata is maintained in the code repository.

Resource maintainer: shbiao shbiao@stu.nepu.edu.cn.

The public resource retains component-specific provenance. COCO annotations, original images, CLASH captions, SugarCrepe material, and newly derived annotation/results have distinct rights. The license record identifies upstream terms and components without an explicit redistribution or reuse license. Public access does not grant additional rights to third-party material. A blanket MIT license is not applied to the Dataset.