Monday 22nd April: 13.00 – 17.00

In this workshop, you will learn practical techniques for customizing LLMs for quant finance using prompt engineering, retrieval augmentation, and fine-tuning.

Prior knowledge of LLMs or Python programming not required. Open-source examples will be provided for those interested in running and modifying the code (CPU or GPU).

The workshop will include two 90-minute sessions (13:00 – 14:30 and 15:00 – 16:30) with 30 min coffee and Q&A breaks after each session.

Models: GPT-3.5, GPT-4, Llama 2, Code Llama

Session One: Prompting and Retrieval Augmentation – 13:30 to 14:30

• Prompting – natural language programming of LLMs
  • Principles of prompt engineering
  • Prompt types
• Retrieval augmentation – using information outside model training
  • Embedding – asking questions over documents
  • Chains – multi-step workflows
  • Memory
• Overcoming limitations
  • Context window
  • Large documents
  • Hallucinations
  • Reproducibility
• Performance Optimization
  • CPU and GPU performance profiles
  • Quantization
• Hands-on examples with Python
  • Comprehension of trade confirmations

Q&A – 14:30 to 15:00

Session Two: Fine-Tuning – 15:00 to 16:30

• Unsupervised fine-tuning
  • Expanding the model dataset and vocabulary
• Self-supervised fine-tuning
  • Reducing prompt length
• Supervised fine-tuning
  • Generated datasets
  • Curated datasets
• Performance Optimization
  • LORA and QLORA
• Hands-on examples with Python
  • Generation of draft model release notes

Q&A – 16:30 to 17:00