List your company
ResearchFormerly WizWand

Multimodal Benchmarking on MMBench (Language Specific Scores)

77MMBench Score (en)DPVR-PC
45.2853.51561.7569.985Nov 8, 2024Feb 12, 2025May 19, 2025Aug 23, 2025Nov 27, 2025Mar 3, 2026Jun 8, 2026
Updated 4mo ago

Evaluation results

MethodLinks
DPVR-PCModel Scale=13B, Split layer (s)=282026.067774
DPVR-PCModel Scale=13B, Split layer (s)=202026.0676.973.8
DPVR-PCModel Scale=13B, Split layer (s)=242026.0676.873.9
DPVR-LFModel Scale=13B, Split layer (s)=282026.0676.673.9
DPVR-PCModel Scale=13B, Split layer (s)=342026.0676.573.9
LLaVA-JPEG-901+Fine-tuned=Yes (1 additional epoch), Input=JPEG images, JPEG quality=902024.1167.0957.3
LLaVA-OriFine-tuned=No, Input=Original images2024.1166.6758.93
LLaVA-JPEG-90Fine-tuned=No, Input=JPEG images, JPEG quality=902024.1166.3258.59
LLaVA-Ori1+Fine-tuned=Yes (1 additional epoch), Input=Original images2024.1165.6457.56
LLaVA-AlignedVQFine-tuned=No, Input=AlignedVQ features2024.1165.3756.7
VisNecSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0364.959.1
PreSelSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0364.856.5
Full-Data(665K)Sampling Ratio=100% (665K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0364.358.3
TypiClustSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0364.357.1
XMASSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0364.354.1
LLaVA-JPEG-10Fine-tuned=No, Input=JPEG images, JPEG quality=102024.1164.1754.46
CoIDOSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0363.355.5
ICONSSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0363.155.8
LLaVA-JPEG-101+Fine-tuned=Yes (1 additional epoch), Input=JPEG images, JPEG quality=102024.1163.0554.2
RandomSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.036153.5
COINCIDESampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0360.553.9
EL2NSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0358.248.5
IFDSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0357.250.6
Full-FinetuneDataset=Vision-Flan-186K, #Data=186k2026.0552.645.9
MAGICDataset=Vision-Flan-186K, #Data=∼37k2026.055246.5
CLIP-ScoreSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.035148
Self-FilterSampling Ratio=15% (98K samples), Base Model=LLaVA-v1.5-7B, Training Dataset=LLaVA-665K2026.0348.145.4
RandomDataset=Vision-Flan-186K, #Data=∼37k2026.0546.539.6