轩辕系列大模型是度小满金融领域的大型语言模型，旨在为金融行业提供强大的中文对话和推理能力。该系列模型包括从6B到176B不同参数规模的版本，并针对金融场景进行了深度优化，能够处理金融事件解读、业务分析、投研应用、合规风控等任务。

How It Works

轩辕系列模型，特别是XuanYuan-FinX1，采用了创新的“思维链+过程奖励+强化学习”训练范式，旨在提升模型在复杂金融决策分析中的逻辑推理能力。通过构建稳定的思维链生成模型、金融决策加强的双奖励模型（结果导向ORM和过程级PRM），并结合PPO算法进行强化学习微调，模型能够生成透明的思考过程并给出更准确的答案。XuanYuan3-70B系列则基于LLaMA3-70B，通过增量预训练和RLHF对齐，在金融场景下表现媲美GPT-4o。

Quick Start & Requirements

• Installation: Typically via Hugging Face transformers library.
• Dependencies: Python, PyTorch, transformers, accelerate, bitsandbytes (for quantization), vllm (for inference acceleration). Specific CUDA versions may be required for optimal performance.
• Hardware: Varies by model size; 70B models generally require multiple 80GB GPUs, while 4-bit quantized versions can run on a single 40GB GPU.
• Resources: Official documentation and GitHub repositories provide detailed usage examples and setup guides.

Highlighted Details

• Financial Domain Expertise: Outperforms leading open-source models and rivals GPT-4o on financial benchmarks like FinanceIQ.
• Reasoning Capabilities: XuanYuan-FinX1 demonstrates advanced logical reasoning and transparent thought processes.
• Model Variety: Offers a range of model sizes (6B, 13B, 70B, 176B) and quantized versions (8-bit, 4-bit) to suit different hardware constraints.
• Long Context Support: Models like XuanYuan3-70B and XuanYuan2-70B support up to 16k context length.

Maintenance & Community

The project is actively developed by Duxiaoman AI-Lab, with regular releases of new models and updates. Links to Hugging Face, ModelScope, and WeChat news are provided for community engagement and updates.

Licensing & Compatibility

The models are generally released under permissive licenses allowing for research and commercial use, but specific terms should be verified for each model version.

Limitations & Caveats

While benchmark results are strong, actual performance may vary in real-world applications. The project is continuously evolving, and users should refer to the latest documentation for compatibility and usage details.