README.md

---
pretty_name: ViRel-Budget
language:
- en
license: other
task_categories:
- visual-question-answering
tags:
- vision-language-models
- visual-token-pruning
- energy-efficiency
- green-ai
- multimodal-evaluation
- reliability
- acm-mm
- greenmm
size_categories:
- 1K<n<10K
---

# ViRel-Budget

**Reliability-constrained visual-token budgeting for green vision-language inference**

Accepted for an oral presentation at the ACM MM 2026 GreenMM Workshop.

[![DOI](https://img.shields.io/badge/DOI-10.5281%2Fzenodo.22014989-blue.svg)](https://doi.org/10.5281/zenodo.22014989)

- **Code:** [StableTradeAtlas/ViRel-Budget](https://github.com/StableTradeAtlas/ViRel-Budget)
- **Archival release:** [Zenodo record 22014989](https://zenodo.org/records/22014989)
- **Dataset:** [StableTradeAtlas/ViRel-Budget](https://huggingface.co/datasets/StableTradeAtlas/ViRel-Budget)

## Dataset summary

ViRel-Budget releases evaluation manifests, derived annotations, configuration files, and documentation for studying reliability-constrained visual-token pruning in vision-language inference. It supports separate evaluation of task correctness, dense-answer agreement, intervention-defined response preservation, GPU energy, latency, memory, and reliability-adjusted efficiency.

The public evaluation design contains a 1,200-query development pool and 900 group-isolated prospective queries. The principal systems comparison is additionally evaluated on three independently sampled, group-disjoint 210-case workloads (630 cases total). These repeated workloads test robustness of the systems result; they are not presented as independent replications of the full study.

> ViRel-Budget measures **intervention-defined visual reliance**. It does not certify universal semantic grounding or whether a model used the correct visual evidence. Intervention outputs supervise and evaluate the controller but are unavailable to it at deployment.

## Supported uses

This release is intended for:

- reproducing the reported evaluation protocol and saved-result analyses;
- auditing task correctness separately from intervention-defined behavior preservation;
- comparing dense inference with FastV, SCOPE, and a deterministic Random control under the documented configurations;
- studying whether measured device-energy savings remain positive after reliability qualification and repeated-workload evaluation.

It is not intended as a general-purpose VQA training corpus, a universal grounding benchmark, or evidence that token reduction necessarily reduces device energy.

## Data sources and configurations

The evaluation manifests reference public source datasets rather than redistributing their images:

| Source | Canonical identifier | Configuration / split used |
|---|---|---|
| MMStar | [Lin-Chen/MMStar](https://huggingface.co/datasets/Lin-Chen/MMStar) | validation split |
| POPE | [lmms-lab/POPE](https://huggingface.co/datasets/lmms-lab/POPE) | Full; adversarial, popular, and random subsets |
| Visual CounterFact | [mgolov/Visual-Counterfact](https://huggingface.co/datasets/mgolov/Visual-Counterfact) | default; color and size subsets |

Users must obtain source images from the original providers and comply with each source dataset's license and terms.

## Evaluation design and leakage controls

Prospective grouping uses the source-image SHA-256 together with normalized exact non-template question frequency below five. This produces 819 groups among the 900 prospective cases, with no source-image overlap between development and prospective partitions.

Cases for which an intervention has no eligible visual tokens are excluded from that intervention's safety denominator rather than counted as vacuously safe. Eligible prospective counts are:

| Backend | SCOPE | FastV | Random |
|---|---:|---:|---:|
| LLaVA-1.5-7B | 296 / 900 | 297 / 900 | 297 / 900 |
| LLaVA-1.5-13B | 253 / 900 | 251 / 900 | 251 / 900 |

Prospective reliability results and repeated-workload energy results belong to different evaluation populations and should not be pooled.

## Main systems result

On the separate 630-query repeated-workload population, SCOPE produced positive measured GPU-energy reductions on all three draws: 11.87% for the 7B backend (95% CI 7.19%–16.30%) and 10.62% for the 13B backend (95% CI 7.48%–13.46%). Under the evaluated implementations, FastV and Random did not show consistent device-energy savings. These findings are bounded to the documented models, workloads, software revisions, and NVIDIA RTX PRO 6000 Blackwell Server Edition measurement environment.

## Reproduction

Use the canonical code repository for scripts, pinned external revisions, test fixtures, measurement boundaries, and reproduction instructions:

**https://github.com/StableTradeAtlas/ViRel-Budget**

The public code release pins the FastV, SCOPE, and LLaVA-PruMerge dependencies used by the study and records CodeCarbon 3.2.9 for prospective carbon accounting. Saved-result inspection does not require a GPU; model-backed execution requires the documented external environments and model weights.

## Limitations and responsible use

- The controller is model-, pruner-, workload-, and hardware-specific.
- Task/source indicators are strong controller features, so transfer to unseen task families is not established.
- Intervention-defined preservation is a behavioral criterion, not proof of causal or semantically correct grounding.
- Carbon values are operational estimates based on measured GPU energy and disclosed PUE/grid assumptions, not lifecycle assessments.
- Source images, model weights, full historical intermediate grids, private mappings, and raw telemetry streams are not redistributed.
- Derived annotations can inherit selection, label, language, and measurement biases from the source datasets and evaluated models.
- The released manifests are not designed to contain personal or sensitive information, but users should consult and follow the source datasets' documentation.

## Open data and code statement

To support transparency, reproducibility, and reuse, we release evaluation manifests, derived annotations, configuration files, and accompanying documentation through this Hugging Face dataset. The [canonical GitHub repository](https://github.com/StableTradeAtlas/ViRel-Budget) releases the source code, evaluation scripts, summarized experimental outputs, and compact raw repeated-workload records. Source images, model weights, complete historical intermediate grids, and private mappings are not redistributed; users must obtain external materials from their original providers under the applicable licenses and terms.

## Citation

Please cite the accompanying camera-ready paper and the archived release. Machine-readable citation metadata is provided in this dataset repository's [`CITATION.cff`](CITATION.cff). The canonical code repository also maintains [software citation metadata](https://github.com/StableTradeAtlas/ViRel-Budget/blob/main/CITATION.cff).

DOI: [10.5281/zenodo.22014989](https://doi.org/10.5281/zenodo.22014989)

## Licensing

The GitHub code is licensed under Apache-2.0. This dataset repository contains derived manifests and metadata whose reuse remains subject to the licenses and terms of the referenced source datasets. The `license: other` metadata value is intentional and avoids implying that one blanket license overrides upstream rights.