weaverbirdllm/famma

Introduction

FAMMA is a multi-modal financial Q&A benchmark dataset. The questions encompass three heterogeneous image types - tables, charts and text & math screenshots - and span eight subfields in finance, comprehensively covering topics across major asset classes. Additionally, all the questions are categorized by three difficulty levels — easy, medium, and hard - and are available in three languages — English, Chinese, and French. Furthermore, the questions are divided into two types: multiple-choice and open questions.

More importantly, FAMMA provides a "live" benchmark for evaluating financial analysis capabilities of LLMs. The benchmark continuously collects new questions from real-world financial professionals, ensuring up-to-date and contamination-free evaluation.

The leaderboard is regularly updated and can be accessed at https://famma-bench.github.io/famma/.

The project code is available at https://github.com/famma-bench/bench-script.

NEWS

🔥 Latest Updates:

• [2025/02] Release of release_v2501 dataset.
• [2025/01] Release of release_v2406 dataset, now including answers and explanations with enhanced quality.
• [2024/06] Initial public release of FAMMA benchmark (based on the release_v2406 dataset), along with our paper: FAMMA: A Benchmark for Financial Domain Multilingual Multimodal Question Answering.

Dataset Versions

FAMMA is continuously updated with new questions. We provide different versions of the dataset:

• release_v2406: The release containing 1935 questions, collected from online sources. Apart from the questions, both answers and explanations are provided.
• release_v2501: The release containing 100 questions, created by invited experts. Only the questions are provided.

Dataset Structure

• idx：a unique identifier for the index of the question in the dataset.
• question_id: a unique identifier for the question across the whole dataset: {language}{main_question_id}{sub_question_id}_{release_version}.
• context: relevant background information related to the question.
• question: the specific query being asked.
• options: the specific query being asked.
• image_1- image_7: directories of images referenced in the context or question.
• image_type: type of the image, e.g., chart, table, screenshot.
• answers: a concise and accurate response. (public on release v2406, non-public on the live set release v2501)
• explanation：a detailed justification for the answer. (public on release v2406, non-public on the live set release v2501)
• topic_difficulty: a measure of the question's complexity based on the level of reasoning required.
• question_type: categorized as either multiple-choice or open-ended.
• subfield: the specific area of expertise to which the question belongs, categorized into eight subfields.
• language：the language in which the question text is written.
• main_question_id：a unique identifier for the question within its context; questions with the same context share the same ID.
• sub_question_id：a unique identifier for the question within its corresponding main question.
• ans_image_1 - ans_image_6： (public on release v2406, non-public on the live set release v2501)

Download

see the script at https://github.com/famma-bench/bench-script/blob/main/step_1_download_dataset.py

Fristly, clone the repository and install the dependencies:

git clone https://github.com/famma-bench/bench-script.git
pip install -r requirements.txt

To download the dataset, run the following command:

python step_1_download_dataset.py \
    --hf_dir "weaverbirdllm/famma" \
    --split "release_v2406" \ # or "release_v2501" or None to download the live set
    --save_dir "./hf_data"

Options:

• --hf_dir: HuggingFace repository name
• --split: Specific version to download (optional)
• --save_dir: Local directory to save the dataset (default: "./hf_data")

Citation

If you use FAMMA in your research, please cite our paper as follows:

@article{xue2024famma,
  title={FAMMA: A Benchmark for Financial Domain Multilingual Multimodal Question Answering},
  author={Siqiao Xue, Tingting Chen, Fan Zhou, Qingyang Dai, Zhixuan Chu, and Hongyuan Mei},
  journal={arXiv preprint arXiv:2410.04526},
  year={2024},
  url={https://arxiv.org/abs/2410.04526}
}