Complete anti-fraud validation leaderboard

README.md

@@ -141,6 +141,18 @@ vision_zh = load_dataset("MTSAIR/MWS-Vision-Bench", "zh")
 
 ---
 
+## Update — July 23, 2026
+
+### New: experimental anti-fraud category
+
+Anti-fraud is reported as a standalone research metric and is not included in
+`Overall`, which remains the mean of the five original benchmark categories.
+The public validation data is available in
+[`MTSAIR/MWS-Antifraud-Bench`](https://huggingface.co/datasets/MTSAIR/MWS-Antifraud-Bench)
+under its own dataset license.
+
+---
+
 ## Update — February 16, 2026
 
 ### New: VQA category update
@@ -221,13 +233,13 @@ Each line in `metadata.jsonl` contains one JSON object:
 
 Top models evaluated on this validation dataset:
 
-`Overall` is the mean of the five original categories. `Anti-fraud` is reported separately and is not included in `Overall`. `—` means that no official anti-fraud evaluation is available.
+`Overall` is the mean of the five original categories. `Anti-fraud` is reported separately and is not included in `Overall`.
 
 | Model | Overall | img→text | img→markdown | Grounding | KIE (JSON) | VQA | Anti-fraud |
 |-------|---------|----------|--------------|-----------|------------|-----|------------|
-| Claude Fable 5 | 0.799 | 0.828 | 0.737 | 0.635 | 0.843 | 0.950 | — |
+| Claude Fable 5 | 0.799 | 0.828 | 0.737 | 0.635 | 0.843 | 0.950 | 0.521 |
 | GPT-5.6 Sol | 0.790 | 0.818 | 0.716 | 0.621 | 0.867 | 0.926 | 0.467 |
-| Kimi K3 | 0.777 | 0.783 | 0.716 | 0.588 | 0.856 | 0.941 | — |
+| Kimi K3 | 0.777 | 0.783 | 0.716 | 0.588 | 0.856 | 0.941 | 0.400 |
 | GPT-5.6 Terra | 0.749 | 0.808 | 0.693 | 0.571 | 0.799 | 0.874 | 0.404 |
 | Kimi K2.6 | 0.743 | 0.868 | 0.761 | 0.301 | 0.858 | 0.929 | 0.279 |
 | Cotype Pro 3 | 0.736 | 0.861 | 0.747 | 0.384 | 0.832 | 0.858 | 0.233 |
@@ -240,7 +252,7 @@ Top models evaluated on this validation dataset:
 | Gemini-2.5-pro | 0.690 | 0.840 | 0.717 | 0.070 | 0.888 | 0.935 | 0.088 |
 | Kimi K2 Instruct | 0.686 | 0.874 | 0.726 | 0.148 | 0.805 | 0.876 | 0.331 |
 | Cotype Light 3 | 0.685 | 0.855 | 0.770 | 0.125 | 0.873 | 0.799 | 0.318 |
-| GPT-5.4 | 0.683 | 0.764 | 0.627 | 0.351 | 0.826 | 0.846 | — |
+| GPT-5.4 | 0.683 | 0.764 | 0.627 | 0.351 | 0.826 | 0.846 | 0.432 |
 | Claude Sonnet 4.6 | 0.682 | 0.827 | 0.725 | 0.186 | 0.755 | 0.915 | 0.235 |
 | Gemini-3-flash-preview | 0.681 | 0.836 | 0.724 | 0.051 | 0.845 | 0.950 | 0.153 |
 | Gemini-3.1-flash-lite-preview | 0.674 | 0.846 | 0.741 | 0.047 | 0.813 | 0.923 | 0.217 |
@@ -252,7 +264,7 @@ Top models evaluated on this validation dataset:
 | GPT-5.2 | 0.663 | 0.799 | 0.656 | 0.173 | 0.855 | 0.835 | 0.413 |
 | Alice AI VLM dev | 0.662 | 0.881 | 0.777 | 0.063 | 0.747 | 0.841 | 0.046 |
 | GPT-4.1-mini | 0.659 | 0.863 | 0.735 | 0.093 | 0.750 | 0.853 | 0.173 |
-| Qwen3.5-27B | 0.658 | 0.773 | 0.699 | 0.078 | 0.856 | 0.882 | 0.188 |
+| Qwen3.5-27B | 0.658 | 0.773 | 0.699 | 0.078 | 0.856 | 0.882 | 0.190 |
 | Cotype VL (32B 8 bit) | 0.649 | 0.802 | 0.754 | 0.267 | 0.683 | 0.737 | 0.078 |
 | GPT-5-mini | 0.639 | 0.782 | 0.678 | 0.117 | 0.774 | 0.843 | 0.233 |
 | Qwen3.5-9B | 0.625 | 0.766 | 0.649 | 0.075 | 0.782 | 0.852 | 0.178 |
@@ -267,7 +279,6 @@ Top models evaluated on this validation dataset:
 | GPT-4.1 | 0.574 | 0.692 | 0.681 | 0.093 | 0.624 | 0.779 | 0.131 |
 | Mistral Large 3 2512 | 0.565 | 0.777 | 0.713 | 0.065 | 0.560 | 0.709 | 0.215 |
 | Mistral Small 3.2 24B Instruct | 0.561 | 0.734 | 0.695 | 0.060 | 0.599 | 0.715 | 0.122 |
-| Qwen3-VL-4B-Instruct | 0.515 | 0.699 | 0.702 | 0.061 | 0.506 | 0.607 | — |
 | Qwen3.5-2B | 0.489 | 0.743 | 0.621 | 0.041 | 0.466 | 0.574 | 0.277 |
 | Qwen3.5-0.8B | 0.319 | 0.549 | 0.499 | 0.027 | 0.157 | 0.362 | 0.185 |