LLMs and Generative AI

To Do

• LLM Configuration
• LLM In-context learning (ICL)
• LLM Quantization
• Three LLMs comparison
• Instruction fine-tuning (IFT): Single-task
• Instruction fine-tuning (IFT): Multi-task
• Parameter-efficient fine-tuning (PEFT)
• Reinforcement learning with human feedback (RLHF)
• LLM Powered Applications

Generative AI Project Lifecycle

A Generative AI project can be broadly divided into three major stages:

1. Scope & Model Selection

• Clearly define the business problem or use case before starting development.
• Identify success metrics and constraints (cost, latency, accuracy, compliance).
• Model selection criteria: 1. Pretraining alignment: How well the model’s pretraining data and objectives align with your domain and task. 2. Model size: Trade-offs between performance, inference latency, and computational resources. 3. Context window: Maximum input length supported by the model, which impacts document handling, reasoning depth, and memory.

2. Adaptation & Alignment

• Prompt engineering
• Fine-tuning or parameter-efficient tuning (e.g., LoRA)
• Retrieval-Augmented Generation (RAG)
• Safety, bias, and alignment adjustments
• The goal is to improve task performance while ensuring reliability and responsible behavior.

3. Application Integration

• API or service integration
• Frontend and user experience design
• Monitoring, logging, and evaluation
• Cost optimization and scalability
• Deployment and maintenance