The LLM Pro Finance Suite: Multilingual Large Language Models for Financial Applications

Battle Count: 87

The paper is highly relevant to quantitative trading infrastructure as it provides domain-specialized LLMs for financial text processing, regulatory document analysis, financial translation, and RAG-based financial Q&A. These capabilities directly support quantitative trading workflows including: parsing regulatory filings for alpha signals, multilingual market analysis, automated financial reporting interpretation, and risk assessment from unstructured text. However, the paper does not directly address trading strategy development, price prediction, or portfolio optimization. The models serve as NLP infrastructure components rather than direct trading signal generators.

Training requires substantial computational resources (up to 256 Nvidia H100 GPUs for the 70B model, 10,000 GPU hours). The data curation pipeline involves multi-stage filtering with a 235B parameter LLM judge, custom classifier training, and synthetic data generation. However, the released 8B models can be used as drop-in replacements with standard inference infrastructure. Fine-tuning on custom data would require DeepSpeed ZeRO-3, TRL library, and access to H100-class GPUs. The evaluation pipeline requires reformatting benchmarks to chat templates and implementing custom LLM-as-a-Judge prompts.

The paper provides detailed training hyperparameters (batch size, gradient accumulation, learning rate, optimizer), compute resources, and dataset composition. Two 8B models are publicly released on HuggingFace. However, the full training dataset is not released, some evaluation benchmarks are in-house/private, and the data curation pipeline details (specific web crawling sources, classifier training) are described but not fully reproducible without access to proprietary data (AGEFI articles, in-house parallel translation data). The LLM-as-a-Judge prompts and metrics formulas are provided in appendices.