OTC-Bench: Overlay–Target Confusion Benchmark

Version 0.1.0-rc1. Research preview with frozen evaluation records.

Code and evaluation · Dataset and cached outputs

Associated manuscript in preparation: When Overlays Become Targets: Diagnosing Target–Distractor Confusion in Vision-Language Models. No paper or DOI is claimed yet.

This diagnostic benchmark asks whether a vision-language model names an artificial object-pattern distractor as a specified original target. A paired exchange moves identical distractor pixels from the target to a nearby non-overlapping slot, while retaining the same visible target pixels. The benchmark measures directed donor-category answers, target accuracy, other errors and UNKNOWN separately.

Left: all 1,000 pairs from 500 target images. Right: model-specific conditioned subsets, with different denominators. Donor-answer counts include the existing format-tolerant scoring; strict category answers and descriptions are separated in the released sensitivity table. See main results, scoring, and figure captions.

Contents and counts

• 800 distinct COCO val2014 target images, one annotated target per image.
• 16 shared donor crops from eight categories; two different donor categories per target.
• 1,600 target-donor pairs; 6,400 unique input images, not 6,400 independent scenes.
• Six cached model runs, 6,400 answers each: 38,400 formal answers.
• Earlier cohorts: 96 + 204 targets. Later frozen validation: 500 targets. Report them separately before any descriptive 800-target aggregate.
• Human audit: 3,900 judgments on 1,300 target-donor pairs. Earlier 300: one donor per target, three readings. Later 500: both donors per target, three readings. Do not claim 4,800 human judgments or full two-donor human coverage on all 800 images.

Conditions

Code	What is on the target?	What is near it?	Purpose
O	Original pixels	Gray slot	Baseline; not an unmodified original photograph
G	Gray rectangle	Gray slot	Gray-occluded baseline
T	Object-pattern donor	Gray slot	Target-overlap condition
N-near	Gray rectangle	Identical object-pattern donor	Matched spatial exchange
donor-check	Donor at its T size/location	Entire remaining canvas gray	Check donor recognition at rendered size

The original image is on the left; a gray sidebar and gray sidebar slot remain in every condition. A four-pixel red target-box outline specifies the original target. T/N-near do not restore hidden target pixels. Position, visual attention and scene interpretation can all change; the pair is not proof of an internal mechanism.

Quick start

Clone either repository to obtain the complete portable evaluation package:

git clone https://github.com/Seren666/OTC-Bench.git
cd OTC-Bench

Use Python 3.11 or newer in a fresh environment. On Windows:

python -m venv .venv
.venv/Scripts/python -m pip install -r requirements.txt
.venv/Scripts/python benchmark.py verify
.venv/Scripts/python benchmark.py evaluate --predictions predictions/idefics3.jsonl --output my_results.csv

On Linux/macOS replace .venv/Scripts/python with .venv/bin/python. Evaluation itself uses only the Python standard library. This command reproduces cached answers; it does not call a model.

To reconstruct inputs, first review LICENSE_STATUS.md and the source-image licenses:

.venv/Scripts/python benchmark.py download-sources --sources sources
.venv/Scripts/python benchmark.py render --sources sources --output images

Downloading source images is an explicit user action; evaluation never downloads anything. If the COCO host is unavailable, obtain the official COCO val2014 archive and place the listed original JPEGs in sources/. SHA256 checks reject modified files. Rendering requires the libjpeg-turbo Pillow build: a non-turbo JPEG decoder can differ by one intensity value and will correctly fail the frozen RGB-hash check. Do not bypass the check or compare new images with old model answers.

Evaluate another model

Read only data/model_inputs.jsonl in the inference adapter; it contains neutral IDs, image paths and the exact prompt. Do not pass category labels, answers, human votes or semantic filenames to the model. Apply that model's native preprocessing/chat template and document its revision, precision, image splitting and deterministic decoding. Never retry a valid but incorrect answer to improve it.

Write one JSON object per line:

{"request_id":"T0001__O","answer":"oven"}

Provide exactly all 6,400 IDs, or use --cohort new_500 with exactly its 4,000 IDs. Blank, missing and duplicate answers fail validation; they are engineering failures, not UNKNOWN. A model's explicit UNKNOWN is a valid recorded response.

.venv/Scripts/python benchmark.py evaluate --predictions my_model.jsonl --scoring format_tolerant --output my_model_results.csv
.venv/Scripts/python benchmark.py evaluate --predictions my_model.jsonl --scoring frozen --output my_model_frozen.csv

The default format_tolerant rule is the already-used uniform alias-in-sentence score, not a new judge model. frozen preserves the historical short-category scorer. Read SCORING.md: neither parser is a complete semantic adjudicator, and donor mentions must not automatically be called strict identity substitutions.

Files

• data/targets.jsonl: source IDs, labels, segmentation, fixed R/P/S rectangles, sizes, cohort and donor assignments.
• data/donors.jsonl, sources.jsonl: crop recipes, original hashes, COCO/Flickr URLs and recorded licenses.
• data/requests.jsonl: full evaluation keys, exact prompts and frozen RGB/PNG hashes.
• data/model_inputs.jsonl: label-free model-facing input list.
• data/human_votes.jsonl: aggregate frozen vote counts, without names, timestamps, clicks or free-text reasons.
• predictions/: cached raw answer strings under neutral request IDs.
• configs/: recorded model revisions and runtime parameters; missing historical revision metadata is explicit.
• results/: cohort-separated scores, all-pair and conditioned scopes, donor strata, both scoring modes.
• figures/: quantitative plots and captions. Photo examples are retained in the owner's internal review bundle pending rights review.

Use and limitations

This is a stress-test diagnostic for artificial object-pattern occlusion, not a natural-occlusion prevalence estimate. Shared donors limit inference across donor instances; confidence intervals condition on this donor bank. Model-specific reliability filters produce different denominators, so use full common-cohort results for comparisons and conditional rows to diagnose failure, not to claim an adjusted leaderboard.

The new 500 targets were frozen before their outputs were inspected but were drawn from familiar COCO imagery and use the existing donor bank. Do not describe them as universally unseen by models or as an untouched donor holdout. No train/test split for method development is established here. Reserve new independent evaluation if later tuning a repair method.

Human prompts were simplified Chinese rather than literally identical to the model's English prompt. One reader knew the research; repeat exposure to two donor versions may matter. Human votes support a readability analysis, not a definitive human-vs-model ability ranking. See the three microwave-target scorer-coverage exception in SCORING.md.

Citation and release status

The project owner has authorized public distribution at the two repositories above. Cite OTC-Bench by repository URL and the exact commit used until a paper and archival identifier are available. No paper DOI, dataset DOI or final author list has been assigned. Cite COCO as the source dataset as well. Component-specific rights and the current code/metadata licensing status are recorded in LICENSE_STATUS.md; public accessibility does not override source-image terms.