naghamo/mist-vlm-judges

MIST - Misleading-Image Stroop Test for VLM Judges

Anonymous release accompanying a double-blind submission. 200 potentially idiomatic English compounds annotated by two disjoint panels of three human annotators each, plus labels produced by 13 vision-language models under 4 prompting techniques, 3 image conditions, and 2 instruction variants.

Source and construction

We build on a public instruction-tuning release, UCSC-Admire/idiom-SFT-dataset-561, which extends the AdMIRe data: 551 English potentially idiomatic compounds, each with an idiomatic row and a literal row - the same compound used figuratively in one context sentence and literally in the other, each row shipping a model-generated image of that reading.

We sample 200 compounds, one per row, 50 per condition. The sentence sense is the sense the compound carries in its sentence. Crossing it with the sense the shipped image depicts gives four conditions:

Condition	Sentence sense	Shipped image sense	Item type
T1	idiom	literal	misleading
T2	idiom	idiom	aligned
T3	literal	literal	aligned
T4	literal	idiom	misleading

An item is aligned when sentence and image senses agree (T2, T3), misleading when they do not (T1, T4). Pairing a sentence with the other row's image yields the misleading half without generating or curating anything.

Two vocabularies: aligned / misleading vs matched / mismatched

• Alignment is a property of the item: it is what an annotator saw, and it is fixed per row.
• Matching is a property of an input to a judge: an image is matched when it depicts the sentence sense, mismatched when it depicts the other. Because judges see all three inputs (text-only, shipped image, swapped image), matching is defined for every input on every row - unlike alignment, it is not tied to what the annotator saw.

The two are one fact from two sides: an item's shipped image is matched when the item is aligned, and mismatched when the item is misleading.

Two images per row

Every row carries both images of the compound:

• image - the image shown to the human annotators for this row. Its sense is image_type.
• swapped_image - the opposite-sense image (never shown to any annotator). Its sense is swapped_image_type.

For VLM judges we ran each row three times (image conditions):

• original = row's image
• swapped = row's swapped_image
• none = no image (text only)

Selecting the matched / mismatched image

By the aligned-item rule above:

• alignment == 'aligned' -> matched = image, mismatched = swapped_image
• alignment == 'misleading' -> matched = swapped_image, mismatched = image

or, equivalently, using the type columns (works without checking alignment):

from datasets import load_dataset

ds = load_dataset('naghamo/mist-vlm-judges', split='full')

def matched_and_mismatched(row):
    """Return (matched_image, mismatched_image) for one row.
    matched  = image whose sense matches the sentence sense
    mismatched = image whose sense is opposite to the sentence sense
    """
    if row['image_type'] == row['sentence_usage']:
        return row['image'], row['swapped_image']
    else:
        return row['swapped_image'], row['image']

matched, mismatched = matched_and_mismatched(ds[0])

So every item has three inputs available for judges: (text alone), (text, matched image), (text, mismatched image), and each of the 260 vlm_... columns is the label a specific model produced under one of those inputs (see the naming below).

Structure

• 200 rows, one per (compound, alignment) item, delivered as a single HF split named full (no train/val/test).
• Two panels (share no items): informed (items 1-100, annotators 1-3); blind (items 101-200, annotators 4-6).
• Label space (partially ordinal): Fully Figurative (FF), Weak Figurative (WF), Figurative and Literal (FL), Fully Literal (LL).

Files

• HF parquet (via datasets.load_dataset) - 200 items with the two typed Image columns (image, swapped_image) and all tabular fields.
• dataset.csv - the same 200 items as a plain CSV, minus the image bytes (image column contents are in the parquet only).
• README.md / schema.json - dataset card and full column dictionary.

Column groups

Item identity: example_id, panel, compound, text, type, sentence_usage, alignment.

Images (typed as datasets.Image in the parquet): image, swapped_image. Their senses are recorded in image_type and swapped_image_type.

Human labels: annotator1_label..annotator3_label on the informed panel; annotator4_label..annotator6_label on the blind panel. Each label is one of the four values above; take a per-item majority if you need a single reference label per item.

VLM labels: 260 columns of the form vlm_<model>__<prompt>__<image_condition>__<instruction>. The <image_condition> is one of none / original / swapped and maps to the three inputs above.

Notes on use

• Fewshot exclusion: for any analysis using the few or few_cot prompts, drop example_id in {13, 46, 51, 159} - those items are the fewshot demonstrations shown in-context.
• Panel identity is meaningful. The paper treats the informed and blind panels as separate references; do not pool them without a reason.

License

CC-BY-4.0.

Anonymization

Anonymous release for double-blind review. Annotators are identified only by number; no author names, affiliations, timestamps, or contact details are embedded in the data.