MWS-Vision-Bench

🇷🇺 Русскоязычное описание ниже / Russian summary below.

MWS Vision Bench — the first Russian-language business-OCR benchmark designed for multimodal large language models (MLLMs).
This is the validation split - publicly available for open evaluation and comparison.
🧩 Paper is coming soon.

Language Configs

This dataset provides three Hugging Face configs for the same benchmark split:

• default — Russian questions
• en — English questions
• zh — Chinese questions

from datasets import load_dataset

vision_ru = load_dataset("MTSAIR/MWS-Vision-Bench")
vision_en = load_dataset("MTSAIR/MWS-Vision-Bench", "en")
vision_zh = load_dataset("MTSAIR/MWS-Vision-Bench", "zh")

🔗 Official repository: github.com/mts-ai/MWS-Vision-Bench
🏢 Organization: MTSAIR on Hugging Face
📰 Article on Habr (in Russian): “MWS Vision Bench — the first Russian business-OCR benchmark”

---

Update — February 16, 2026

New: VQA category update

We updated the Reasoning VQA (ru) category to improve evaluation robustness. Revised questions and answers only (images remain unchanged). Results are not directly comparable to versions prior to Feb 16, 2026 - for VQA and overall columns.

This update improves reliability of reasoning-based evaluation while keeping the benchmark structure intact.

---

📊 Dataset Statistics

• Total samples: 1,302
• Unique images: 400
• Task types: 5

---

🖼️ Dataset Preview

Examples of diverse document types in the benchmark: business documents, handwritten notes, technical drawings, receipts, and more.

---

📁 Repository Structure

MWS-Vision-Bench/
├── metadata.jsonl       # Dataset annotations
├── images/              # Image files organized by category
│   ├── business/
│   │   ├── scans/
│   │   ├── sheets/
│   │   ├── plans/
│   │   └── diagramms/
│   └── personal/
│       ├── hand_documents/
│       ├── hand_notebooks/
│       └── hand_misc/
└── README.md            # This file

---

📋 Data Format

Each line in metadata.jsonl contains one JSON object:

{
  "file_name": "images/image_0.jpg",   # Path to the image
  "id": "1",                           # Unique identifier
  "type": "text grounding ru",         # Task type
  "dataset_name": "business",          # Subdataset name
  "question": "...",                   # Question in Russian
  "answers": ["398", "65", ...]        # List of valid answers (as strings)
}

---

🎯 Task Types

Task	Description	Count
document parsing ru	Parsing structured documents	243
full-page OCR ru	End-to-end OCR on full pages	144
key information extraction ru	Extracting key fields	119
reasoning VQA ru	Visual reasoning in Russian	400
text grounding ru	Text–region alignment	396

---

📊 Leaderboard (Validation Set)

Top models evaluated on this validation dataset:

Model	Overall	img→text	img→markdown	Grounding	KIE (JSON)	VQA
Claude-4.6-Opus	0.704	0.841	0.748	0.168	0.852	0.908
Qwen3.6-27B	0.691	0.863	0.768	0.071	0.833	0.919
Gemini-2.5-pro	0.690	0.840	0.717	0.070	0.888	0.935
Gemini-3-flash-preview	0.681	0.836	0.724	0.051	0.845	0.950
Gemini-3.1-flash-lite-preview	0.674	0.846	0.741	0.047	0.813	0.923
Gemini-2.5-flash	0.672	0.886	0.729	0.042	0.825	0.879
Claude-4.5-Opus	0.670	0.809	0.720	0.131	0.799	0.889
Claude-4.5-Sonnet	0.669	0.741	0.660	0.459	0.727	0.759
GPT-5.2	0.663	0.799	0.656	0.173	0.855	0.835
Alice AI VLM dev	0.662	0.881	0.777	0.063	0.747	0.841
GPT-4.1-mini	0.659	0.863	0.735	0.093	0.750	0.853
Cotype VL (32B 8 bit)	0.649	0.802	0.754	0.267	0.683	0.737
GPT-5-mini	0.639	0.782	0.678	0.117	0.774	0.843
Qwen3-VL-235B-A22B-Instruct	0.623	0.812	0.668	0.050	0.755	0.830
Qwen2.5-VL-72B-Instruct	0.621	0.847	0.706	0.173	0.615	0.765
GPT-5.1	0.588	0.716	0.680	0.092	0.670	0.783
Qwen3-VL-8B-Instruct	0.584	0.780	0.700	0.084	0.592	0.766
Qwen3-VL-32B-Instruct	0.582	0.730	0.631	0.056	0.708	0.784
GPT-4.1	0.574	0.692	0.681	0.093	0.624	0.779
Qwen3-VL-4B-Instruct	0.515	0.699	0.702	0.061	0.506	0.607

Scale: 0.0 - 1.0 (higher is better)

📝 Submit your model: To evaluate on the private test set, contact g.gaikov@mts.ai

---

💻 Usage Example

from datasets import load_dataset

# Load dataset (authorization required if private)
dataset = load_dataset("MTSAIR/MWS-Vision-Bench", token="hf_...")

# Example iteration
for item in dataset:
    print(f"ID: {item['id']}")
    print(f"Type: {item['type']}")
    print(f"Question: {item['question']}")
    print(f"Image: {item['image_path']}")
    print(f"Answers: {item['answers']}")

---

📄 License

MIT License
© 2024 MTS AI

See LICENSE for details.

---

📚 Citation

If you use this dataset in your research, please cite:

@misc{mwsvisionbench2024,
  title={MWS-Vision-Bench: Russian Multimodal OCR Benchmark},
  author={MTS AI Research},
  organization={MTSAIR},
  year={2025},
  url={https://huggingface.co/datasets/MTSAIR/MWS-Vision-Bench},
  note={Paper coming soon}
}

---

🤝 Contacts

• Team: MTSAIR Research
• Email: g.gaikov@mts.ai

---

🇷🇺 Краткое описание

MWS Vision Bench — первый русскоязычный бенчмарк для бизнес-OCR в эпоху мультимодальных моделей.
Он включает 1302 примера и 5 типов задач, отражающих реальные сценарии обработки бизнес-документов и рукописных данных.
Датасет создан для оценки и развития мультимодальных LLM в русскоязычном контексте.
📄 Научная статья в процессе подготовки (paper coming soon).

---

Made with ❤️ by MTS AI Research Team