README.md

---
license: mit
task_categories:
- visual-question-answering
- video-classification
- multiple-choice
language:
- en
tags:
- video-language-models
- metaphorical-understanding
pretty_name: ViMU
---

# ViMU: Benchmarking Video Metaphorical Understanding

ViMU is a benchmark for evaluating video-language models on video metaphorical understanding. The benchmark is designed to test whether models can infer implicit, subtextual, rhetorical, and socially situated meanings in videos, rather than relying only on literal visual perception.

The dataset contains curated videos, task annotations, question-answer pairs, evidence-grounding labels, and taxonomy labels for evaluating model behavior across open-ended and structured understanding tasks.

## Dataset Access

This repository is currently maintained under an anonymous account for review purposes.

During anonymous review, access may be restricted or gated to preserve the integrity of the submission process. Access can be granted to reviewers or organizers when needed.

If the submission is accepted, the dataset will be made available according to the stated license and release policy.

## Dataset Structure

The dataset is organized as follows:

```text
ViMU/
├── videos/
│   ├── <video_id>.mp4
│   └── ...
├── metadata/
│   ├── open_ended_qa.jsonl
│   ├── evidence_grounding.jsonl
│   ├── rhetoric_mechanism_mcq.jsonl
│   ├── social_value_signal_mcq.jsonl
│   └── ...
├── dataset.croissant.json
├── LICENSE
└── README.md
````

The exact file names may vary depending on the released version.

## Tasks

ViMU includes four main evaluation tasks.

### 1. Open-Ended Question Answering

The open-ended task evaluates whether a model can explain the implicit meaning of a video in free-form natural language.

Each example includes:

```json
{
  "video_id": "...",
  "question": "...",
  "answer": "..."
}
```

### 2. Evidence Grounding

The evidence-grounding task evaluates whether a model can identify which evidence sources support the intended meaning of the video.

Typical evidence types include:

* visual frames
* visible text
* transcript or speech content
* audio tone
* editing pattern

This task is designed to test whether a model grounds its interpretation in observable video evidence rather than producing unsupported explanations.

### 3. Rhetorical Mechanism Understanding

This task evaluates whether a model can identify the rhetorical mechanism through which the video conveys its implicit meaning.

Examples of rhetorical mechanisms include irony, sarcasm, exaggeration, contrast, parody, deadpan delivery, role reversal, innuendo, and bait-and-switch. For stable evaluation, fine-grained labels are grouped into higher-level macro categories.

### 4. Social Value Signal Understanding

This task evaluates whether a model can identify the social stance, attitude, or value signal conveyed by the video.

Examples of social value signals include contempt, norm violation, aggression or hostility, anti-mainstream sentiment, political or identity signaling, and sexual implication. For stable evaluation, fine-grained labels are grouped into higher-level macro categories.

## Data Fields

Common fields may include:

| Field                     | Description                                                      |
| ------------------------- | ---------------------------------------------------------------- |
| `video_id`                | Unique identifier for the video                                  |
| `video_path`              | Relative path to the video file                                  |
| `question`                | Evaluation question                                              |
| `answer`                  | Reference answer for open-ended QA                               |
| `options`                 | Multiple-choice options, when applicable                         |
| `gold_label`              | Ground-truth answer or label                                     |
| `gold_options`            | Ground-truth option IDs for multi-select tasks                   |
| `rhetoric_mechanisms`     | Fine-grained rhetorical mechanism labels                         |
| `social_value_signals`    | Fine-grained social value signal labels                          |
| `target_subjects`         | Target or subject category, when applicable                      |
| `evidence_sources`        | Evidence modalities supporting the interpretation                |
| `difficulty`              | Estimated difficulty level                                       |
| `cultural_dependency`     | Estimated dependence on cultural or platform-specific context    |
| `interpretability_bucket` | Whether the video is self-contained or requires external context |

## Example Record

```json
{
  "video_id": "example_video_id",
  "question": "What implicit meaning does the video convey?",
  "answer": "The video uses contrast and exaggeration to imply a critical or humorous interpretation beyond the literal scene.",
  "evidence_sources": ["frames", "visible_text", "editing_pattern"],
  "rhetoric_mechanisms": ["contrast", "exaggeration"],
  "social_value_signals": ["norm_violation"],
  "difficulty": "medium",
  "cultural_dependency": "medium"
}
```

This example is illustrative. Please refer to the released metadata files for the exact schema.

## Intended Uses

ViMU is intended for:

* evaluating video-language models on implicit and metaphorical video understanding;
* benchmarking model performance beyond literal visual recognition;
* analyzing evidence grounding in video interpretation;
* studying model behavior on rhetorical mechanisms and social value signals;
* supporting reproducible academic research on multimodal reasoning.

## Out-of-Scope Uses

The dataset is not intended for:

* identifying or profiling individuals appearing in videos;
* making automated moderation decisions without human review;
* training systems for surveillance, targeting, or personal attribute inference;
* commercial deployment without permission;
* making broad claims about cultural competence or social reasoning without careful analysis.

## Responsible Use and Limitations

ViMU contains videos and annotations involving implicit meanings, social cues, rhetorical devices, and potentially sensitive social signals. The benchmark is designed for evaluation and research, not for identifying individuals or inferring private attributes.

Important limitations include:

* The dataset may overrepresent internet-native or platform-specific video conventions.
* Some examples may require cultural, linguistic, or contextual knowledge.
* The benchmark should not be interpreted as a complete measure of general video intelligence.
* Model performance may depend on frame sampling, video preprocessing, prompt design, and decoding settings.
* Taxonomy labels may involve subjective judgment despite validation and filtering.

Users should report results with appropriate caveats and avoid overgeneralizing from aggregate scores.

## Biases

The dataset may contain biases from its source distribution, curation process, and annotation taxonomy. Potential biases include:

* overrepresentation of certain meme formats or online communities;
* underrepresentation of some cultures, languages, identities, or rhetorical traditions;
* ambiguity in socially situated or culturally dependent interpretations;
* possible annotation subjectivity for implicit meaning and social value signals.

Researchers are encouraged to analyze performance across difficulty levels, cultural-dependency levels, evidence types, and taxonomy categories.

## Personal and Sensitive Information

Some videos may contain publicly visible people, social signals, identity-related references, political or social commentary, or other sensitive implications.

The dataset should not be used to identify people, infer private attributes, or make decisions about individuals or groups.

## Croissant Metadata

This dataset includes Croissant metadata for NeurIPS dataset submission and machine-readable dataset documentation.

The Croissant file is provided at:

```text
dataset.croissant.json
```

It describes the dataset distribution, metadata files, record sets, and Responsible AI fields.

## License

This dataset is released under a Research Use Only License.

The dataset is provided solely for non-commercial research purposes. Users may use, download, and analyze the dataset for academic and research activities, including model evaluation, benchmarking, and reproducibility studies.

Commercial use, redistribution for commercial purposes, or use in products, services, or systems intended for commercial deployment is not permitted without prior written permission from the dataset maintainers.

By using this dataset, users agree to use it responsibly, with appropriate consideration of privacy, fairness, and potential social impact.

See the `LICENSE` file for the full license terms.

## Citation

The citation will be updated after the review process.

For anonymous review, please cite as:

```bibtex
@misc{anonymous2026vimu,
  title        = {ViMU: Benchmarking Video Metaphorical Understanding},
  author       = {Anonymous},
  year         = {2026},
  note         = {Under review}
}
```

## Contact

For anonymous review, please use the official review system or the contact mechanism provided by the hosting platform.