Datasets:

SeanWan05
/

ViRel-Budget

dataset_stratum stringclasses 6 values	gpu_energy_joule float64 11 397	latency_ms float64 79.6 1.67k	method stringclasses 3 values	model_scale stringclasses 2 values	sample_id stringclasses 900 values	selected_budget stringclasses 4 values	supported_correct bool 2 classes	task_correct bool 2 classes	unsafe_acceptance bool 2 classes
mmstar	20.558293	101.592122	fastv	7b	mmstar_000011	dense	false	false	false
mmstar	22.958695	99.106692	fastv	7b	mmstar_000012	dense	false	false	false
mmstar	63.554998	236.677604	fastv	7b	mmstar_000028	dense	false	true	false
mmstar	28.084747	94.712252	fastv	7b	mmstar_000031	dense	true	true	false
mmstar	33.334625	97.256733	fastv	7b	mmstar_000037	dense	false	true	false
mmstar	74.392878	208.201503	fastv	7b	mmstar_000043	dense	false	true	false
mmstar	63.807483	176.8049	fastv	7b	mmstar_000049	432	false	true	false
mmstar	59.54088	168.625329	fastv	7b	mmstar_000054	dense	false	false	false
mmstar	33.447773	95.67125	fastv	7b	mmstar_000068	dense	true	true	false
mmstar	60.254753	170.596423	fastv	7b	mmstar_000089	dense	false	true	false
mmstar	41.676328	118.136626	fastv	7b	mmstar_000091	dense	false	true	false
mmstar	32.438518	92.114095	fastv	7b	mmstar_000092	dense	false	true	false
mmstar	35.118833	101.01605	fastv	7b	mmstar_000094	dense	true	true	false
mmstar	33.092324	95.537381	fastv	7b	mmstar_000103	dense	true	true	false
mmstar	32.404416	90.773255	fastv	7b	mmstar_000105	dense	false	false	false
mmstar	35.540955	98.812463	fastv	7b	mmstar_000113	432	false	false	false
mmstar	46.85009	124.855703	fastv	7b	mmstar_000124	dense	false	false	false
mmstar	35.788545	94.565881	fastv	7b	mmstar_000131	dense	true	true	false
mmstar	36.065563	95.071305	fastv	7b	mmstar_000139	dense	true	true	false
mmstar	35.541104	93.664255	fastv	7b	mmstar_000154	432	false	false	false
mmstar	35.594222	93.662674	fastv	7b	mmstar_000155	dense	true	true	false
mmstar	35.484563	93.346061	fastv	7b	mmstar_000157	432	false	false	false
mmstar	37.369802	96.677024	fastv	7b	mmstar_000160	dense	true	true	false
mmstar	35.394017	91.539128	fastv	7b	mmstar_000163	dense	true	true	false
mmstar	36.90656	95.47586	fastv	7b	mmstar_000166	432	false	true	false
mmstar	36.525888	94.493149	fastv	7b	mmstar_000167	432	false	true	false
mmstar	35.079077	91.413487	fastv	7b	mmstar_000174	432	true	true	false
mmstar	34.824491	90.838723	fastv	7b	mmstar_000183	dense	true	true	false
mmstar	40.811093	106.899083	fastv	7b	mmstar_000184	432	false	true	false
mmstar	34.701027	90.996925	fastv	7b	mmstar_000187	dense	false	true	false
mmstar	35.287344	92.623996	fastv	7b	mmstar_000192	dense	false	true	false
mmstar	35.070374	92.067283	fastv	7b	mmstar_000193	dense	true	true	false
mmstar	35.038806	92.002774	fastv	7b	mmstar_000196	dense	true	true	false
mmstar	37.911558	99.534959	fastv	7b	mmstar_000197	432	false	false	false
mmstar	75.152936	195.551729	fastv	7b	mmstar_000205	dense	false	false	false
mmstar	36.060996	93.286803	fastv	7b	mmstar_000207	432	false	true	false
mmstar	35.889326	92.76576	fastv	7b	mmstar_000219	dense	true	true	false
mmstar	89.789176	233.067429	fastv	7b	mmstar_000243	dense	false	true	false
mmstar	46.316203	120.531252	fastv	7b	mmstar_000244	dense	true	true	false
mmstar	34.519466	89.962026	fastv	7b	mmstar_000248	dense	true	true	false
mmstar	36.276899	94.913201	fastv	7b	mmstar_000254	432	false	false	false
mmstar	39.499027	103.748668	fastv	7b	mmstar_000257	432	false	true	false
mmstar	35.154975	93.017798	fastv	7b	mmstar_000267	432	true	true	false
mmstar	117.642036	337.718315	fastv	7b	mmstar_000270	dense	true	true	false
mmstar	29.026329	92.147883	fastv	7b	mmstar_000274	432	false	false	false
mmstar	53.252665	172.97093	fastv	7b	mmstar_000275	dense	false	false	false
mmstar	31.24379	101.155167	fastv	7b	mmstar_000283	432	false	false	false
mmstar	45.680934	146.976588	fastv	7b	mmstar_000284	432	false	true	false
mmstar	30.0842	98.144298	fastv	7b	mmstar_000291	dense	false	false	false
mmstar	28.974411	94.356932	fastv	7b	mmstar_000311	dense	false	false	false
mmstar	30.393136	93.166463	fastv	7b	mmstar_000313	432	false	false	false
mmstar	31.820105	97.442482	fastv	7b	mmstar_000320	432	false	false	false
mmstar	35.991645	98.990284	fastv	7b	mmstar_000324	432	false	false	false
mmstar	33.816767	92.985	fastv	7b	mmstar_000325	432	true	true	false
mmstar	34.163043	91.493057	fastv	7b	mmstar_000335	432	false	false	false
mmstar	34.721969	93.025381	fastv	7b	mmstar_000344	dense	false	false	false
mmstar	39.456987	105.772916	fastv	7b	mmstar_000354	dense	false	true	false
mmstar	35.091462	94.008248	fastv	7b	mmstar_000360	432	false	true	false
mmstar	36.518207	96.477415	fastv	7b	mmstar_000366	432	false	false	false
mmstar	50.730211	134.576168	fastv	7b	mmstar_000367	432	false	false	false
mmstar	38.094627	101.930054	fastv	7b	mmstar_000369	dense	false	false	false
mmstar	37.268909	101.128377	fastv	7b	mmstar_000370	dense	true	true	false
mmstar	36.28991	99.674787	fastv	7b	mmstar_000375	432	true	true	false
mmstar	32.940593	91.649687	fastv	7b	mmstar_000379	dense	false	false	false
mmstar	32.400893	90.691613	fastv	7b	mmstar_000384	432	false	false	false
mmstar	32.269343	90.881374	fastv	7b	mmstar_000386	432	false	false	true
mmstar	34.308374	96.871808	fastv	7b	mmstar_000397	432	false	false	true
mmstar	33.216281	94.256697	fastv	7b	mmstar_000400	dense	false	false	false
mmstar	61.823898	175.110428	fastv	7b	mmstar_000405	dense	false	false	false
mmstar	36.867944	103.201384	fastv	7b	mmstar_000407	432	false	false	false
mmstar	36.161668	101.066647	fastv	7b	mmstar_000413	dense	false	false	false
mmstar	55.972855	157.620037	fastv	7b	mmstar_000423	432	false	true	false
mmstar	36.106568	104.051201	fastv	7b	mmstar_000427	432	false	false	false
mmstar	31.954975	93.15509	fastv	7b	mmstar_000428	432	false	false	false
mmstar	31.478799	92.349427	fastv	7b	mmstar_000431	dense	false	false	false
mmstar	30.876736	91.013873	fastv	7b	mmstar_000433	432	false	false	false
mmstar	33.913298	100.181913	fastv	7b	mmstar_000438	432	false	false	false
mmstar	31.502065	92.458606	fastv	7b	mmstar_000440	dense	true	true	false
mmstar	31.026219	90.826053	fastv	7b	mmstar_000444	432	false	false	false
mmstar	36.345949	105.083087	fastv	7b	mmstar_000447	432	true	true	false
mmstar	55.420105	159.06278	fastv	7b	mmstar_000453	432	false	false	false
mmstar	36.811182	104.747968	fastv	7b	mmstar_000464	dense	false	false	false
mmstar	35.539077	100.688545	fastv	7b	mmstar_000466	dense	false	false	false
mmstar	35.617225	100.467144	fastv	7b	mmstar_000470	432	false	false	false
mmstar	32.05931	90.619752	fastv	7b	mmstar_000473	432	false	false	true
mmstar	49.50904	140.168989	fastv	7b	mmstar_000479	dense	false	false	false
mmstar	33.926182	95.89299	fastv	7b	mmstar_000481	dense	true	true	false
mmstar	33.357702	94.264947	fastv	7b	mmstar_000485	dense	false	false	false
mmstar	33.667715	94.927382	fastv	7b	mmstar_000491	432	false	true	false
mmstar	42.776583	120.146442	fastv	7b	mmstar_000494	432	false	false	false
mmstar	35.857787	99.928251	fastv	7b	mmstar_000495	432	false	false	false
mmstar	54.296895	150.083952	fastv	7b	mmstar_000500	dense	false	true	false
mmstar	33.014106	92.702128	fastv	7b	mmstar_000501	dense	false	false	false
mmstar	32.788102	93.462795	fastv	7b	mmstar_000508	dense	false	false	false
mmstar	34.009965	96.917048	fastv	7b	mmstar_000509	432	false	false	false
mmstar	85.912846	244.897524	fastv	7b	mmstar_000511	432	false	false	false
mmstar	38.147378	111.261392	fastv	7b	mmstar_000512	dense	false	false	false
mmstar	31.263218	91.839589	fastv	7b	mmstar_000519	dense	false	false	false
mmstar	32.548117	95.95122	fastv	7b	mmstar_000521	432	false	false	false
mmstar	33.146385	95.417688	fastv	7b	mmstar_000538	432	false	false	true

End of preview. Expand in Data Studio

ViRel-Budget metadata and derived evaluation artifacts

This release accompanies “ViRel-Budget: Reliability-Constrained Visual-Token Budgeting for Green Vision-Language Inference,” accepted for an oral presentation at the ACM Multimedia 2026 GreenMM workshop.

Authors: Sean Wan, Shilin Ou, and Luyao Zhang.

Release boundary

The package contains the frozen query-selection manifests, group identifiers, budget-specific behavioral-fidelity labels, prospective controller decisions, and three-draw replication membership/results used by the paper.

It does not contain source images, questions, answers, or answer options. Those materials remain with MMStar, POPE, and Visual CounterFact and must be obtained from their official providers. The query manifests provide source repository, configuration, split, and row identifiers for reconstruction.

Configurations

query_manifests: 1,200 development and 900 group-isolated prospective selections.
safety_labels: derived labels for two LLaVA-1.5 scales, FastV/SCOPE/Random, and four pruned budgets.
prospective_decisions: one-call decisions and execution outcomes for the six frozen prospective controllers.
replication_membership: three fixed, group-disjoint draws of 210 unique queries.
replication_decisions: matched one-call decisions and measured energy/latency outcomes for the replication population.

Loading

from datasets import load_dataset

queries = load_dataset("SeanWan05/ViRel-Budget", "query_manifests")
labels = load_dataset("SeanWan05/ViRel-Budget", "safety_labels")
prospective = load_dataset("SeanWan05/ViRel-Budget", "prospective_decisions")
replications = load_dataset("SeanWan05/ViRel-Budget", "replication_decisions")

GitHub repository: https://github.com/SeanWan514/ViRel-Budget HuggingFace repository: https://huggingface.co/datasets/SeanWan05/ViRel-Budget

Safety-label interpretation

reference_safe means that a pruned action preserved the dense model's original answer and its responses under all eligible interventions. It is an operational behavioral-fidelity label, not proof of semantic grounding or correct causal evidence use. Task correctness is reported separately where applicable.

The 1,200-query development labels were used for controller fitting/calibration. The 900-query prospective labels were hidden until all controller decisions had been recorded.

Energy and carbon boundary

GPU energy is measured within the declared call window. Carbon values in the associated paper are estimates derived from measured energy plus disclosed grid-intensity and PUE assumptions; they are not direct emissions measurements or lifecycle assessments.

Limitations

The controller evidence covers LLaVA-1.5 7B and 13B only. The selected tasks come from three benchmark families. Intervention-defined fidelity may reflect uncertainty, distribution shift, or intervention artifacts, and it inherits errors from the dense reference. Results do not establish transfer to unrelated VLM backbones, hardware, or modalities.

Source datasets and licensing

The original ViRel-Budget metadata and derived labels/results in this repository are released under the Apache License 2.0; see LICENSE. See LICENSES.md for the third-party boundary. Users must retrieve source content from the official upstream repositories and comply with their current terms.

Integrity

hf_source_inventory.json records source paths, checksums, and the code-repository commit used for packaging.
validation_report.json records structural and leakage checks.
release_manifest.json records the checksum and size of every upload candidate file.
dataset_schema.json documents the configurations and primary keys.

Citation

Final ACM citation metadata will be added after the DOI and proceedings metadata are issued.

Acknowledgments

Sean Wan and Shilin Ou gratefully acknowledge support from the Summer Research Scholars Program at Duke Kunshan University, under the supervision of Prof. Luyao Zhang, and from the Duke Kunshan University Library grant supporting the Open Data Contest project.

Downloads last month: 53