Image-Text Retrieval on Retrieval
74.78Avg RecallBaseline
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BaselineBackbone=ViT-H/14-quickgelu2026.01 | 74.78 | — | |
| TuneCLIPBackbone=ViT-H/14-quickgelu2026.01 | 73.78 | — | |
| OursModel=Qwen2.5-VL2026.04 | 64.03 | 95.8 | |
| Grid-SearchModel=Qwen2.5-VL2026.04 | 63.11 | 95 | |
| Fixed-AmpModel=Qwen2.5-VL2026.04 | 62.91 | 94.92 | |
| Ours w/o KLModel=Qwen2.5-VL2026.04 | 62.81 | 94.75 | |
| BaseModel=Qwen2.5-VL2026.04 | 62.59 | 94.64 | |
| RandNeuronModel=Qwen2.5-VL2026.04 | 62.53 | 94.8 | |
| OursModel=LLaVA-1.52026.04 | 59.46 | 96.26 | |
| Grid-SearchModel=LLaVA-1.52026.04 | 58.7 | 95.6 | |
| Ours w/o KLModel=LLaVA-1.52026.04 | 57.8 | 93.65 | |
| Fixed-AmpModel=LLaVA-1.52026.04 | 57.73 | 93.7 | |
| BaseModel=LLaVA-1.52026.04 | 57.23 | 93.28 | |
| RandNeuronModel=LLaVA-1.52026.04 | 57.18 | 93.45 | |
| MLMType=D3, Dataset Size=38M2024.05 | 29 | — | |
| DFN ∪ Ours (20%)*Type=D3, Dataset Size=23M2024.05 | 28.1 | — | |
| DFN ∪ HYPEType=D3, Dataset Size=20M2024.05 | 28 | — | |
| DFN ∪ HYPE ∪ Ours (10%)*Type=D3, Dataset Size=22M2024.05 | 27.6 | — | |
| DFNType=D2, Dataset Size=16M2024.05 | 27 | — | |
| Ours (20%)Type=D1, Dataset Size=22M2024.05 | 26.3 | — | |
| T-MARSType=D3, Dataset Size=22M2024.05 | 25.4 | — | |
| DevilType=D3, Dataset Size=20M2024.05 | 24.7 | — | |
| HYPEType=D3, Dataset Size=10M2024.05 | 22.2 | — | |
| OAI CLIP Filtered data + AllBackbone=CLIP ViT-B/32, Training Samples=12.8M, Training Steps=4 x 12.8M, Evaluation Mode=Zero-shot2026.06 | 14.9 | — | |
| Self-Filtered CLIP + AllBackbone=CLIP ViT-B/32, Training Samples=12.8M, Training Steps=4 x 12.8M, Evaluation Mode=Zero-shot2026.06 | 14.8 | — | |
| All dataBackbone=CLIP ViT-B/32, Training Samples=12.8M, Training Steps=4 x 12.8M, Evaluation Mode=Zero-shot2026.06 | 14.7 | — |