Multimodal Instruction Tuning on LLaVA-665K
100.2Relative ScoreVisNec
Evaluation Results
| Method | Links | |
|---|---|---|
| VisNecSelection Time=12.0 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=×, Total Cost=23.0 GPU-hr2026.03 | 100.2 | |
| Full FinetuneSelection Time=–, Finetuning Time=76.0 GPU-hr, API Cost=×, Total Cost=76.0 GPU-hr2026.03 | 100 | |
| PreSelSelection Time=9.0 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=✓, Total Cost=20.0 GPU-hr + API Cost2026.03 | 97.7 | |
| CoidoSelection Time=25.0 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=✓, Total Cost=36.0 GPU-hr + API Cost2026.03 | 97.7 | |
| XMASSelection Time=36.0 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=×, Total Cost=47.0 GPU-hr2026.03 | 97.3 | |
| COINCIDESelection Time=55.5 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=×, Total Cost=66.5 GPU-hr2026.03 | 95.8 | |
| Self-FilterSelection Time=73.5 GPU-hr, Finetuning Time=11.0 GPU-hr, API Cost=×, Total Cost=84.5 GPU-hr2026.03 | 89.3 |