PIXIU provides a comprehensive suite of resources for developing, fine-tuning, and evaluating financial Large Language Models (LLMs). It includes the FinMA LLM, the FIT instruction dataset, and the FinBen evaluation benchmark, aiming to advance open-source financial AI research. The project is suitable for researchers and developers working on LLMs for financial applications.

How It Works

PIXIU offers three core components: FinBen, a benchmark for evaluating financial LLMs across understanding and prediction tasks; FIT, a multi-task, multi-modal instruction dataset for fine-tuning LLMs on financial tasks; and FinMA, the project's own financial LLM, available in 7B and 30B parameter versions. The approach emphasizes diversity, including financial prediction tasks alongside traditional NLP tasks, and supports multiple data modalities like text and time-series.

Quick Start & Requirements

• Installation: Clone the repository (git clone https://github.com/The-FinAI/PIXIU.git --recursive), install requirements (pip install -r requirements.txt), and then install the evaluation package (cd src/financial-evaluation && pip install -e .[multilingual]). A Docker image is also available (sudo docker pull tothemoon/pixiu:latest).
• Prerequisites: BART checkpoint for evaluation (src/metrics/BARTScore/bart_score.pth).
• Links: Pixiu Paper, FinBen Leaderboard.

Highlighted Details

• Offers a broad benchmark (FinBen) covering 30+ financial tasks including sentiment analysis, NER, QA, and stock movement prediction.
• Provides the FIT dataset with 136K instruction samples across NLP and prediction tasks, supporting multi-modal data.
• Releases FinMA v0.1 models (7B and 30B parameters) fine-tuned on LLaMA, available on Huggingface.
• Includes FinMem, an LLM-based agent framework for financial decision-making and trading.

Maintenance & Community

• The project has multiple academic affiliations and recent updates in early 2024.
• A Discord server is available for community interaction: Discord.

Licensing & Compatibility

• The core PIXIU project is licensed under MIT.
• Specific datasets within FinBen have varying licenses (e.g., CC BY-SA 3.0, MIT, Public), requiring careful review for commercial use.

Limitations & Caveats

• The disclaimer strongly warns that the repository is for academic/educational purposes only and does not constitute financial advice, with users assuming all risks.
• Automated evaluation for tasks like NER might yield lower performance in zero-shot settings compared to human annotations.