Tampered Text Classification on TFR CTM Cross-Method
88.8AccuracyTextShield-R1
Evaluation results
| Method | Links | |
|---|---|---|
| TextShield-R1Fine-tuning=Full training set images2026.02 | 88.8 | |
| Qwen2.5-VL-7BFine-tuning=Full training set images2026.02 | 73.6 | |
| FakeShield*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 73.6 | |
| SIDA*Fine-tuning=Full training set images, Base MLLM=Qwen2.5-VL-7B2026.02 | 73.6 | |
| InternVL3-8BFine-tuning=Full training set images2026.02 | 67.7 | |
| MiniCPM_V_2.6Fine-tuning=Full training set images2026.02 | 64.8 | |
| Qwen2.5-VL-3BFine-tuning=Full training set images2026.02 | 63 | |
| InternVL3-2BFine-tuning=Full training set images2026.02 | 62.5 | |
| FakeShieldFine-tuning=Full training set images2026.02 | 57.5 | |
| SIDAFine-tuning=Full training set images2026.02 | 57.5 | |
| Qwen2.5-VL-3BFine-tuning=None2026.02 | 42.2 | |
| GPT4oFine-tuning=None2026.02 | 37.8 | |
| Qwen2.5-VL-7BFine-tuning=None2026.02 | 34 | |
| MiniCPM_V_2.6Fine-tuning=None2026.02 | 26.2 | |
| InternVL3-8BFine-tuning=None2026.02 | 25.2 | |
| InternVL3-2BFine-tuning=None2026.02 | 20.3 |