DISC-FinLLM is a Chinese financial large language model designed to provide professional, intelligent, and comprehensive financial consulting services. It targets users needing specialized financial advice, analysis, and calculations, offering a multi-expert system for diverse financial scenarios.

How It Works

DISC-FinLLM is built upon the Baichuan-13B-Chat base model and fine-tuned using the custom DISC-Fin-SFT dataset, comprising ~250k financial-specific instructions. It employs a multi-expert approach, with four LoRA-tuned modules specializing in financial consulting, text analysis, financial calculations (integrating tools like calculators and equation solvers), and retrieval-augmented Q&A for market analysis. This modular design allows for efficient switching between functionalities without reloading the entire model.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Inference requires PyTorch and Hugging Face Transformers.
• Model weights are available on Hugging Face.
• Demo available at: https://fin.fudan-disc.com
• Technical report: http://arxiv.org/abs/2310.15205

Highlighted Details

• Achieves competitive performance on financial NLP tasks, human-generated financial questions, and financial calculation benchmarks.
• Integrates four specialized tools for mathematical and financial model computations.
• Offers both LoRA and full parameter fine-tuned model weights.
• Includes a DISC-Fin-Eval Benchmark for comprehensive model evaluation.

Maintenance & Community

• Developed by Fudan University's Data Intelligence and Social Computing Lab (Fudan-DISC).
• Project is based on several open-source projects including Baichuan-13B and Langchain-Chatchat.

Licensing & Compatibility

• Licensed under the Apache License.
• Suitable for commercial use and integration into closed-source applications.

Limitations & Caveats

The model is intended for reference and does not replace professional financial analysts. Users are advised to evaluate its output critically, as it may have limitations not yet overcome by current LLMs. The developers assume no liability for issues arising from its use.