Skip to content

Latest commit

 

History

102 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

VisFinEval: A Chinese Financial Knowledge Benchmark for Visual Language Models

简体中文 | English

Table of Contents

  1. Introduction
  2. Dataset
  3. Results
  4. Usage
  5. Outlook
  6. Contact Us
  7. Citation

Introduction

VisFinEval is a large-scale Chinese benchmark platform designed to systematically evaluate the performance of multimodal large language models (MLLMs) in real-world financial business scenarios. The dataset includes 15,848 annotated question-answer pairs, covering eight typical financial image modalities (such as candlestick charts, financial statements, etc.), and is divided into three progressive difficulty levels: financial knowledge and data analysis, financial analysis and decision support, and financial risk control and asset optimization. We conducted zero-shot testing on 21 cutting-edge MLLMs, and the results showed that although the best-performing model, Qwen-VL-max, achieved an accuracy of 76.3%, surpassing non-specialized human performance, it still lagged significantly behind financial experts by more than 14 percentage points in advanced domain-specific tasks such as multi-step numerical reasoning and business process understanding.

example Back to Top


Datasets

Datasets comparison

This table compares multiple QA datasets across various dimensions, including question types, financial capability assessment, and scale.

Benchmark Name Question Type Multi-level Difficulty Scenario Depth Realistic Environment Simulation Official Seal Recognition Financial Relationship Graph Number of Financial Figure Types Number of Financial Scenarios Total Questions Number of Evaluated Models
Text-based Benchmarks
FinDABench Open-ended ✓ - - - - - 5 2400 40
SuperCLUE-Fin Open-ended ✗ - - - - - 6 1000 11
CFBenchmark Open-ended ✗ - - - - - 8 3917 22
FinEval Multiple-choice + Open-ended ✗ - - - - - 9 8351 19
Multimodal Benchmarks
SEEDBENCH Multiple-choice ✗ ✗ ✗ ✗ - - - 19000 18
MMMU Multiple-choice ✓ ✗ ✗ ✗ ✗ - - 11500 30
FinVQA Open-ended ✗ ✗ ✗ ✗ ✗ 2 2 1025 9
FIN-FACT True/False ✗ ✗ ✗ ✗ ✗ 2 5 3369 4
FAMMA Multiple-choice + Open-ended ✓ ✗ ✗ ✗ ✗ 3 8 1758 4
MME-Finance Open-ended ✓ ✗ ✗ ✗ ✗ 6 11 2274 19
VisFinEval (Ours) Multiple-choice + True/False + Open-ended ✓ ✓ ✓ ✓ ✓ 8 15 15848 21

Task Settings

VisFinEval comprises three scenario categories with 15 business scenario types:

Scenario Types:

  1. Financial Knowledge and Data Analysis (Front Office): FDS, CCA, FIA, FERI, SSSB, FIE, FSR
  2. Financial Analysis and Business Decision (Middle Office): FSA, IAI, IA, FMSA
  3. Financial Risk Control and Asset Optimization (Back Office): FSO, FRPA, FDRI, AAA

Question Type Distribution

Scenario Depth Financial Scenario Questions
Financial Knowledge and Data Analysis Financial Data Statistics 3,655
Candlestick Chart Analysis 1,124
Financial Indicator Assessment 1,160
Financial Entity Relationships Interpretation 919
Stock Selection Strategies Backtesting 719
Financial Information Extraction 924
Financial Seal Recognition 199
All 8,700
Financial Analysis and Decision Support Financial Scenario Analysis 2,040
Industry Analysis and Inference 1,361
Investment Analysis 933
Financial Market Sentiment Analysis 316
All 4,650
Financial Risk Control and Asset Optimization Financial Strategy Optimization 111
Financial Risk and Policy Analysis 181
Financial Data Reasoning and Interpretation 1,839
Asset Allocation Analysis 367
All 2,498
VisFinEval All 15,848

Results

Model Evaluation Results

Model Size Limit FDS CCA FIA FERI SSSB FIE FSR FSA FMSA FSO FRPA FDRI AAA WA
Qwen-VL-max Unknown / 78.8 90.5 87.4 89.2 86.2 90.6 77.9 65.3 83.1 82.3 76.8 49.1 58.2 76.3
Qwen-VL-max-latest Unknown / 76.0 84.5 86.1 87.1 79.3 88.6 84.4 59.6 82.6 82.8 79.3 44.0 52.2 73.8
InternVL3-78B 78B / 71.2 83.5 71.4 86.7 79.5 87.8 87.4 64.3 82.1 80.4 78.7 49.1 52.8 72.5
Doubao-1.5-vision-pro-32k Unknown / 75.6 79.0 84.2 85.5 76.8 91.7 74.4 56.7 80.2 79.8 77.3 30.0 54.5 71.7
InternVL2.5-78B 78B / 73.3 77.9 72.3 84.2 84.0 88.4 82.9 63.3 81.5 80.1 75.2 41.0 53.1 71.5
Qwen2.5-VL-72B 72B / 75.9 77.0 72.8 85.4 81.5 88.3 80.4 57.4 82.4 80.3 74.5 41.4 53.4 71.0
GPT-4o-2024-11-20 Unknown / 72.0 76.8 74.9 81.7 71.8 83.8 83.9 61.9 77.9 78.5 73.2 41.0 40.5 68.5
Step-1o-vision-32k Unknown / 48.9 78.4 80.2 84.1 75.3 88.2 98.0 40.3 78.8 78.6 76.1 39.2 45.2 68.4
Moonshot-V1-32k-vision-preview Unknown / 56.2 82.8 73.4 80.5 73.9 87.6 68.3 61.9 77.7 77.0 72.3 39.2 55.8 68.3
Qwen2.5-VL-7B 7B / 71.4 75.9 69.2 80.9 74.0 85.5 69.9 53.4 79.7 76.5 70.7 37.2 37.6 65.4
InternVL3-8B 8B / 68.2 78.0 62.8 87.0 74.1 84.0 77.4 56.5 76.1 76.8 71.7 29.7 46.2 65.4
Gemini-2.5-pro-exp-03-25 Unknown / 73.6 76.7 72.6 81.0 73.0 89.4 87.4 53.2 72.4 70.8 75.5 28.4 38.0 64.7
Claude-3-7-Sonnet-20250219 Unknown / 70.5 73.4 80.3 71.1 77.5 83.2 34.7 48.0 76.1 75.5 64.0 26.8 50.3 62.9
Qwen2.5-VL-3B 3B / 69.5 81.1 65.9 76.6 73.6 83.4 72.4 50.0 75.4 74.7 66.6 22.9 34.8 62.4
MiniCPM-V-2.6 8B / 61.3 83.5 56.9 76.7 75.2 73.4 80.9 48.3 69.7 70.7 69.1 20.6 35.5 60.1
Llama-3.2-11B-Vision-Instruct 11B / 56.9 40.8 59.3 63.9 62.9 73.1 70.4 45.3 69.7 67.1 63.4 18.0 22.1 50.9
Molmo-7B-D-0924 7B / 60.1 74.8 54.5 62.2 59.1 60.5 42.2 39.7 64.4 62.8 63.4 23.4 31.7 49.8
GLM-4v-Plus-20250111 Unknown Multi-image Limit 73.8 86.6 87.9 87.5 81.2 89.3 72.7 56.5 78.1 74.9 74.6 45.1 54.1 72.0
LLaVA-NEXT-34B 34B Context Window Limit 55.3 79.8 92.3 63.2 87.8 55.0 58.8 54.3 88.2 88.1 66.9 13.1 17.5 56.0
LLaVA-v1.6-Mistral-7B 7B Context Window Limit 54.6 73.4 65.9 62.1 47.4 47.0 62.3 42.3 58.3 56.4 63.7 10.2 16.3 47.8
LLaVA-NEXT-13B 13B Context Window Limit 50.2 64.8 43.9 57.2 62.5 50.2 38.7 34.7 59.2 59.0 52.9 14.7 10.8 43.0

Back to Top


Usage Instructions


1.Project Structure

VisFinEval/
├── data/                                              # Raw datasets
│   ├── figure/                                        # Image assets
│   ├── markdown/                                      # Markdown-formatted data
│   ├── Financial Knowledge and Data Analysis          # Front Office Scenario Type data
│   ├── Financial Analysis and Business Decision       # Middle Office Scenario Type data
│   └── Financial Risk Control and Asset Optimization  # Back Office Scenario Type data
├── logs/                  						       # Runtime logs
├── output/                	              		       # Model outputs
├── scripts/             						       # Execution scripts
│   ├── visfineval.py    						       # test scripts
│   └── run_model.sh      						       # Unified evaluation scripts
├── README.md          						           # Project documentation
└── requirements.txt    					           # Dependency specifications

2.Installation Guide

     pip install -r requirements.txt

3. Run All Evaluation Question Types

  • Using the run_model.sh Script to Run All Evaluation Question Types:

    bash run_model.sh
  • Model Configuration for Evaluation:

    • API Model Configuration:

      Modify the model settings according to the comments:

    # API Model Configuration (Format: API_KEY, API_BASE_URL, Model Name)
    USE_API=true 
    API_MODEL_NAME="your_model_name"
    API_KEY="your_model_key"
    API_BASE_URL="your_url"
    • Local Model Configuration:

      Modify the model settings according to the comments:

    # Local Model Configuration (Format: Model Name, Model Path, Model Type, Additional Parameters)
    USE_API=false
    API_MODEL_NAME="your_model_name"
    MODEL_PATH="your_model_address"

4. Output Results

  • The evaluation results are output in the VisFinEval/output and VisFinEval/logs folders.
  • The output results are organized by question type, with each folder further categorized by model.

🚀 Outlook

The evaluation results of FinGAIA deliver a clear message: for the financial industry, merely enhancing the knowledge base of large language models is no longer sufficient. The key to determining whether these models can become reliable productivity tools lies in enabling them to skillfully and collaboratively use various analytical tools, just as humans do.

We plan to:

  • Open-source the dataset: Release an annotated developer question set containing 150 problems.
  • Maintain a leaderboard: Continuously track and showcase the performance of the world's top AI agents in the financial domain through a leaderboard.

The leaderboard will be regularly updated to include more available agents. Individuals and organizations interested in agent evaluation are welcome to connect and exchange ideas.

📫 Contact Us

We sincerely invite industry peers to jointly explore innovative paradigms for the deep integration of AI and finance, and to collaboratively build a new ecosystem for intelligent finance. Please reach out via email at zhang.liwen@shufe.edu.cn.

📚 Citation

If you have used VisFinEval in your research, please cite our paper.

@article{liu2025visfineval,
  title={VisFinEval: A Scenario-Driven Chinese Multimodal Benchmark for Holistic Financial Understanding},
  author={Liu, Zhaowei and Guo, Xin and Xia, Haotian and Zeng, Lingfeng and Lou, Fangqi and Niu, Jinyi and Li, Mengping and Qi, Qi and Li, Jiahuan and Zhang, Wei and others},
  journal={arXiv preprint arXiv:2508.09641},
  year={2025}
}

Back to Top

About

No description, website, or topics provided.

Resources

Stars

8 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages