ENTRAP-VL

Entrainment Assessment Probe for Vision and Language — a taxonomically structured dataset for probing dual contextual entrainment in vision-language models.

ENTRAP-VL is the instrument accompanying the paper ENTRAP-VL: A Taxonomic Probe for Dual Contextual Entrainment in Vision-Language Models (Goyal, Hossain, Das, Bhutani; 2026). It is designed to let the community investigate whether and how a VLM's output is pulled by auxiliary context in its input, separately for context arriving through the textual channel (Textual Entrainment) and the visual channel (Visual Entrainment). Each phenomenon is probed by a structured taxonomy of context conditions.

This dataset is a diagnostic instrument, not a leaderboard. The taxonomy specifies conditions under which pull is anticipated; it is not a set of tasks a good model should always answer correctly. See Intended use below and the paper's Section 7 for the framing.

---

Citing this dataset

Any use of this dataset in any form (research, publication, blog, benchmark tooling, presentation etc.) must cite our arXiv paper. This is stated up front because it is the single most important obligation attached to the release. The required citation is:

@misc{goyal2026entrapvltaxonomicprobedual,
      title={ENTRAP-VL: A Taxonomic Probe for Dual Contextual Entrainment in Vision-Language Models}, 
      author={Karan Goyal and Afreen Hossain and Debojyoti Das and Vishal Bhutani},
      year={2026},
      eprint={2607.20092},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.20092}, 
}

Paper available at https://arxiv.org/abs/2607.20092.

A machine-readable CITATION.cff is provided at the repository root.

---

At a glance

	Textual Entrainment	Visual Entrainment
Fixed input	Image + textual query	Textual query (world-knowledge answer)
Manipulated context	Textual (injected alongside the image-text query)	Visual (candidate images accompanying the query)
Number of conditions	8	3
Categories	8	7 (humans excluded by design)
Items per category	100	100
Total items	800	700
Grand total		1,500

A single record on the textual side carries one image and a structured contexts object enumerating eight context conditions, each holding three statements. A single record on the visual side carries one textual query and three candidate images, one per visual condition.

---

Repository layout

ENTRAP-VL/
├── README.md                       (this file)
├── LICENSE                         (CC BY 4.0 for annotations; source licenses for images)
├── CITATION.cff                    (machine-readable citation metadata)
├── IMAGE_LICENSES.md               (image provenance and per-source licensing)
├── data/
│   ├── textual/<category>/queries.jsonl
│   ├── textual/<category>/images/
│   ├── visual/<category>/queries.jsonl
│   ├── visual/<category>/images/{relatable_true,relatable_distractor,random_distractor}/
│   └── combined/
│       ├── textual_all.jsonl
│       ├── visual_all.jsonl
│       └── entrap_vl_all.jsonl
└── docs/
    ├── taxonomy.md                 (definitions, design, and worked examples)
    ├── schema.md                   (JSONL field specifications)
    └── data_statement.md           (provenance, ethics, intended use)

Categories: creatures, electronics, fruits, household_objects, humans, monuments, natural_landscapes, vehicles. The humans category is intentionally excluded from the visual split; see the paper's Section 7 and docs/data_statement.md.

---

Quick start

Hugging Face datasets

from datasets import load_dataset

textual = load_dataset(
    "goyalkaraniit/ENTRAP-VL",
    data_files="data/combined/textual_all.jsonl",
    split="train",
)
visual = load_dataset(
    "goyalkaraniit/ENTRAP-VL",
    data_files="data/combined/visual_all.jsonl",
    split="train",
)

print(len(textual), "textual items;", len(visual), "visual items")

Direct JSONL

import json
from pathlib import Path

root = Path("ENTRAP-VL/data")
textual = [json.loads(l) for l in (root / "combined" / "textual_all.jsonl").read_text().splitlines() if l.strip()]
visual  = [json.loads(l) for l in (root / "combined" / "visual_all.jsonl").read_text().splitlines()  if l.strip()]

print(len(textual), "textual items;", len(visual), "visual items")

Loading a single textual example

import json
from pathlib import Path
from PIL import Image

cat_dir = Path("ENTRAP-VL/data/textual/creatures")
record = json.loads(cat_dir.joinpath("queries.jsonl").read_text().splitlines()[0])

image  = Image.open(cat_dir / record["image_path"])
prompt = record["prompt"]
for context_type, context_values in record["contexts"].items():
    # context_values is a list of 3 statements for this condition
    for ctx in context_values:
        full_prompt = f"{ctx}\n{prompt}"
        # Feed (image, full_prompt) to your VLM; compare against ground_truth_short
        # to detect textual entrainment.
        ...

Loading a single visual example

import json
from pathlib import Path
from PIL import Image

cat_dir = Path("ENTRAP-VL/data/visual/creatures")
record = json.loads(cat_dir.joinpath("queries.jsonl").read_text().splitlines()[0])

prompt = record["prompt"]
# The world-knowledge answer to `prompt` is record["ground_truth_short"];
# it is known independently of the image. Pull is measured relative to the
# model's no-image answer, not to relatable_true (which is NOT a baseline —
# see docs/taxonomy.md).
for context_type, image_rel in record["images"].items():
    image = Image.open(cat_dir / image_rel)
    # Feed (prompt, image) to your VLM; compare across all three context_types
    # to detect visual entrainment.
    ...

---

Record schema (summary)

Full schema in docs/schema.md.

Textual records (8 fields):

{
  "category": "creatures",
  "item_id": 1,
  "image_path": "images/001.jpg",
  "prompt": "What is the dog doing?",
  "ground_truth_long": "The dog is running through the grass.",
  "ground_truth_short": "running",
  "contexts": {
    "relatable_true":              ["...", "...", "..."],
    "relatable_contradictory":     ["...", "...", "..."],
    "relatable_distractor":        ["...", "...", "..."],
    "random_distractor":           ["...", "...", "..."],
    "relatable_distractor_short":  ["...", "...", "..."],
    "random_distractor_short":     ["...", "...", "..."],
    "relatable_counterfactual":    ["...", "...", "..."],
    "random_counterfactual":       ["...", "...", "..."]
  },
  "entrainment_type": "textual"
}

Visual records (6 fields):

{
  "category": "creatures",
  "query_id": 1,
  "prompt": "A large gray animal with a trunk and tusks is an",
  "ground_truth_short": "elephant",
  "entrainment_type": "visual",
  "images": {
    "relatable_true":       "images/relatable_true/001.jpg",
    "relatable_distractor": "images/relatable_distractor/001.jpg",
    "random_distractor":    "images/random_distractor/001.jpg"
  }
}

item_id (textual) and query_id (visual) are 1-indexed identifiers unique within each category.

---

Taxonomy at a glance

The taxonomy is the conceptual core of ENTRAP-VL. It is organized around two descriptive axes, both defined relative to the item at hand — the depicted image in the textual stream, and the textual query in the visual stream:

• Association: relatable (associated with the item) vs. random (unassociated).
• Veracity: true / contradictory (false of the depicted scene but possible in the world) / counterfactual (false in the world).

The textual stream instantiates both axes and yields eight conditions. The visual stream instantiates only the association axis (an image is a perceptual referent, not a proposition, so veracity does not apply the same way) and yields three conditions.

All eight textual conditions and the three visual conditions are defined and worked out with examples in docs/taxonomy.md.

---

Intended use

ENTRAP-VL is a diagnostic instrument for probing dual contextual entrainment in VLMs. Reduced pull under distractor conditions indicates a model behaving well by the criterion the instrument measures; large pull indicates the phenomenon the instrument is designed to detect. Neither is a general endorsement or condemnation of the model.

Results should be reported by condition rather than aggregated, alongside the no-context (or no-image) reference measurements, so that entrainment magnitudes are interpretable.

ENTRAP-VL is not intended for:

• Training or fine-tuning models. It is a diagnostic set; training on it would both contaminate it as an evaluation instrument and misrepresent its purpose.
• Measuring overall VLM quality. It measures one specific failure mode.
• Any use involving the people depicted in the humans category beyond the evaluation of model behavior (see docs/data_statement.md).

---

Ethics summary

The humans category contains images of real people from royalty-free repositories. Every query about a person is answered by something directly observable (an action, a visible physical state, or an occupation indicated by visible cues). No query asks the model to identify who a person is, and no query infers protected or non-visible attributes. Where a prompt uses affect-adjacent wording, the answer is the observable action (e.g., "screaming"), not an inferred emotional state.

The humans category is excluded entirely from the visual split, since constructing visual distractors for human subjects would require pairing additional images of identifiable people — a consent and identification concern the dataset does not incur.

Full ethical framing, including scale and locale caveats, is in docs/data_statement.md.

---

License

ENTRAP-VL uses a split license:

• Annotations and documentation — all JSONL content (prompts, ground truths, context statements, identifiers, taxonomy structure) and all documentation under docs/ and at the repository root — are released under CC BY 4.0. Attribution to the authors is required; the required attribution form is the arXiv citation above.
• Images are not relicensed. They were sourced from royalty-free repositories (Unsplash, Pexels, Pixabay) and remain subject to the terms of their respective source licenses, which permit free use and redistribution. ENTRAP-VL does not assert copyright over the images. See IMAGE_LICENSES.md.

All images were manually verified to be free of watermarks prior to release.

See LICENSE for the full text.

---

Contact

Karan Goyal · IIIT Delhi, India · karang@iiitd.ac.in

Issues and corrections: please open an issue on this HuggingFace repository.