UniFinEval

UniFinEval is a unified multimodal benchmark for high-information-density financial environments, covering text, images, and videos. It evaluates multimodal large language models (MLLMs) on fine-grained financial reasoning tasks grounded in real-world financial systems.

Paper: UniFinEval: Towards Unified Evaluation of Financial Multimodal Models across Text, Images and Videos
Code: github.com/aifinlab/UniFinEval

Dataset Summary

Item	Value
Total samples	2,515
Total image files	6,568 (referenced)
Languages	Chinese (2,154), English (361)
Question types	Single-round QA (1,341), Multi-round QA (1,174)
Multi-image samples	311

Financial Scenarios

Scenario	Count
金融风险感知 (Financial Risk Sensing)	738
资产配置分析 (Asset Allocation Analysis)	494
公司基本面推导 (Company Fundamental Reasoning)	480
财务报告审计 (Financial Statement Auditing)	438
产业趋势洞察 (Industry Trend Insights)	365

Image Types

image_type	Count	Description
pure_image	858	Chart/table screenshots
mixed	569	Text + image combined
pure_text	553	Text-only financial documents
splice_distractor	213	Spliced distractor samples
single_video	160	Single-video frame sequences
multi_video	150	Multi-video frame sequences
stacked	11	Stacked layout samples
video_image	1	Video + image hybrid

Difficulty Levels

Level	Count
L1	606
L2	975
L3	440
L4	494

Repository Structure

.
├── README.md
├── unifineval-data.csv       # Metadata (questions, answers, paths)
└── organized_dataset/        # Images organized by type and question_id
    ├── mixed/
    │   └── mixed_130/
    │       └── 0004.png
    ├── pure_image/
    ├── single_video/
    └── ...

Data Fields

Field	Type	Description
question_id	string	Unique question identifier
round	string	Round index for multi-round questions (empty for single-round)
question	string	Question text
answer	string	Reference answer
question_type	string	问答题 (single-round) or 多轮问答题 (multi-round)
image_type	string	Modality/layout category (see table above)
image_path	string	Relative path to image(s) from repo root; multiple paths separated by ;
profile	string	Evaluator profile (expert)
scenario	string	Financial scenario category
difficulty	string	Difficulty level (L1–L4)
language	string	cn or en

Usage

Load metadata

import pandas as pd
from pathlib import Path

repo_root = Path(".")  # or downloaded HF repo path
df = pd.read_csv(repo_root / "unifineval-data.csv")
print(df.head())

Resolve image paths

import re
from pathlib import Path

def split_image_paths(raw: str) -> list[str]:
    if not raw or not str(raw).strip():
        return []
    return [p.strip() for p in re.split(r"\s*;\s*|\s*\|\s*", str(raw).strip()) if p.strip()]

repo_root = Path(".")
row = df.iloc[0]
image_files = [repo_root / p for p in split_image_paths(row["image_path"])]
for img in image_files:
    assert img.is_file(), img

Download from Hugging Face Hub

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="AIFin-Lab/UniFinEval",
    repo_type="dataset",
)

Evaluation Notes

• Multi-image rows: 311 samples contain multiple images in image_path (semicolon-separated). Feed all frames/images to the model when evaluating video-derived tasks.
• Multi-round rows: Questions sharing the same question_id with different round values belong to the same multi-turn conversation.
• Zero-shot / CoT: See the paper and GitHub repository for evaluation protocols.

Limitations

• This release contains 2,515 curated samples with verified local image paths. The full benchmark described in the paper includes additional QA pairs across modalities.
• Images are derived from financial research reports, earnings materials, and video frames. Users should comply with applicable copyright and usage policies.

Citation

If you use UniFinEval, please cite:

@article{unifineval2025,
  title   = {UniFinEval: Towards Unified Evaluation of Financial Multimodal Models across Text, Images and Videos},
  author  = {},  % fill in from paper
  journal = {},  % fill in from paper
  year    = {2025}
}

License

This dataset is released under CC BY 4.0.
Please replace the license above if your institution requires a different one.

Upload (maintainers)

From the project root:

export HF_TOKEN=hf_xxx   # never commit this token

# Default: try Tsinghua mirror, auto-fallback to hf-mirror.com if unavailable
bash upload_to_hf.sh

# Or use hf-mirror directly (recommended in China)
bash upload_to_hf.sh --mirror hf-mirror

Dry run (validate paths only):

python3 upload_to_hf.py --dry-run

Note: Tsinghua TUNA (mirrors.tuna.tsinghua.edu.cn/huggingface) no longer provides a working Hugging Face mirror (HTTP 404). The upload script will automatically fall back to hf-mirror.com unless --no-fallback is set.