README.md

---
license: other
license_name: mws-antifraud-research-only-1.0
license_link: LICENSE.md
task_categories:
- image-classification
- visual-question-answering
language:
- ru
tags:
- antifraud
- document-verification
- benchmark
- multimodal
pretty_name: MWS Antifraud Bench (Validation)
dataset_info:
- config_name: default
  features:
  - name: id
    dtype: string
  - name: type
    dtype: string
  - name: dataset_name
    dtype: string
  - name: question
    dtype: string
  - name: answers
    list: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 102549007
    num_examples: 209
  download_size: 102260438
  dataset_size: 102549007
- config_name: en
  features:
  - name: id
    dtype: string
  - name: type
    dtype: string
  - name: dataset_name
    dtype: string
  - name: question
    dtype: string
  - name: answers
    list: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 102449014
    num_examples: 209
  download_size: 102255220
  dataset_size: 102449014
- config_name: zh
  features:
  - name: id
    dtype: string
  - name: type
    dtype: string
  - name: dataset_name
    dtype: string
  - name: question
    dtype: string
  - name: answers
    list: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 102445041
    num_examples: 209
  download_size: 102255620
  dataset_size: 102445041
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
- config_name: en
  data_files:
  - split: train
    path: en/train-*
- config_name: zh
  data_files:
  - split: train
    path: zh/train-*
---

# MWS Antifraud Bench (Validation)

Experimental document-authenticity task for general-purpose multimodal language
models. This is the public validation part of MWS Vision Bench anti-fraud v0.1.

The dataset is released for research and model comparison. It is not a
certification tool, a production fraud-detection system, or a universal
leaderboard that is expected to be resistant to deliberate optimization.

## Data

The validation split contains 209 items:

- 44 `ai_gen`;
- 65 `edited`;
- 100 `original`.

`original` means an unmodified source image in this experiment. It must not be
interpreted as a legal authenticity judgment.

Each item contains:

- `id`;
- `type`;
- `dataset_name` — one of `original`, `edited`, `ai_gen`;
- `question`;
- `answers` — annotated edited fields when available;
- `image`.

## Language configs

The same images are available with three question-language configs:

- `default` — Russian;
- `en` — English;
- `zh` — Chinese.

```python
from datasets import load_dataset

antifraud_ru = load_dataset("MTSAIR/MWS-Antifraud-Bench")
antifraud_en = load_dataset("MTSAIR/MWS-Antifraud-Bench", "en")
antifraud_zh = load_dataset("MTSAIR/MWS-Antifraud-Bench", "zh")
```

## Expected answer

The model must return one JSON object:

```json
{"label": "original|edited|ai_gen", "arguments": "short explanation"}
```

## Metric

The metric combines balanced three-class accuracy and the quality of the
explanation for manually edited documents:

```text
AF = 0.75 × max(0, balanced_accuracy − 1/3)
     + 0.5 × edited_reason_score
```

AF is aggregated over the complete split. Balanced accuracy requires the full
three-class confusion matrix; unlike the five original MWS Vision Bench
metrics, it is not a plain mean of independent per-item scores.

Anti-fraud is reported as a separate leaderboard category and is excluded from
the primary MWS Vision Bench Overall score.

## Leaderboard (Validation)

The table mirrors the `Anti-fraud` column in the
[canonical MWS Vision Bench leaderboard](https://github.com/mts-ai/MWS-Vision-Bench#-leaderboard)
and is sorted by the anti-fraud score. This score is reported separately and
is not included in `Overall`.

| Model | Anti-fraud |
|-------|------------|
| Claude Fable 5 | 0.521 |
| GPT-5.5 | 0.477 |
| GPT-5.6 Sol | 0.467 |
| GPT-5.4 | 0.432 |
| GPT-5.2 | 0.413 |
| GPT-5.6 Terra | 0.404 |
| Kimi K3 | 0.400 |
| Claude Sonnet 5 | 0.391 |
| Claude-4.6-Opus | 0.385 |
| Kimi K2 Instruct | 0.331 |
| Cotype Light 3 | 0.318 |
| Claude-4.5-Opus | 0.308 |
| GPT-5.6 Luna | 0.302 |
| Qwen3-VL-32B-Instruct | 0.298 |
| Claude-4.5-Sonnet | 0.280 |
| Kimi K2.6 | 0.279 |
| Qwen3.5-2B | 0.277 |
| Qwen 3.7 Plus | 0.250 |
| Claude Sonnet 4.6 | 0.235 |
| Cotype Pro 3 | 0.233 |
| GPT-5-mini | 0.233 |
| Qwen3.6-35B-A3B | 0.232 |
| Gemini-3.1-flash-lite-preview | 0.217 |
| Gemini 3.5 Flash | 0.216 |
| Mistral Large 3 2512 | 0.215 |
| Qwen3-VL-8B-Instruct | 0.196 |
| Qwen3.6-27B | 0.195 |
| Qwen3.5-27B | 0.190 |
| Qwen3.5-0.8B | 0.185 |
| GLM-5V Turbo | 0.183 |
| Qwen3.5-9B | 0.178 |
| GPT-4.1-mini | 0.173 |
| Qwen3-VL-235B-A22B-Instruct | 0.171 |
| Gemini-3-flash-preview | 0.153 |
| Qwen3.5-35B-A3B | 0.142 |
| Gemini-2.5-flash | 0.137 |
| GPT-4.1 | 0.131 |
| Mistral Small 3.2 24B Instruct | 0.122 |
| Gemma 4 31B IT | 0.119 |
| Gemini 3.6 Flash | 0.115 |
| Qwen3.5-4B | 0.100 |
| GPT-5.1 | 0.096 |
| Qwen2.5-VL-72B-Instruct | 0.091 |
| Gemini-2.5-pro | 0.088 |
| Cotype VL (32B 8 bit) | 0.078 |
| Alice AI VLM dev | 0.046 |

## Limitations

This is an experimental research dataset. Its class balance and document-type
distribution differ from a production stream and can provide shortcuts to a
system tuned specifically for this collection. A motivated participant can
optimize against the benchmark. Use the score to explore and compare models
for your own scenario, not as a universal measure of document authenticity.

## License and allowed use

This dataset is distributed under the
[MWS Antifraud Research-Only License 1.0](LICENSE.md). The dataset is intended
for research and model comparison and is not a certification or production
fraud-detection system.

The paragraph above is a summary. If it conflicts with `LICENSE.md`, the
license text controls.