Intro to GPT Models

GPT models (Decoders) play a crucial role in generating subsequent words in tasks like text translation or story generation, providing outputs along with their probabilities. They utilize attention mechanisms twice during training: initially, Masked Multi-Head Attention, where only the beginning of a target sentence is revealed, and later, Multi-Head Attention, similar to encoders. In traditional transformer models, decoders interact with encoders by using the encoder’s outputs to assist in tasks like sentence translation. However, GPT models adopt a unique approach by relying solely on a decoder, compensating for the absence of an encoder through extensive training on large datasets. This allows for embedding a vast amount of knowledge within the decoder. ChatGPT further advances these techniques by integrating human-labeled data to address issues such as hate speech and employing Reinforcement Learning for enhanced model quality.

Notebooks - Basics

• Generative Pre-trained Models - Basics
• Generative Pre-trained Models - Basics - Solutions

Notebooks - Applications

• TM Applications - LangChain
• TM Applications - LanceDB
• TM Applications - LanceDB - Solutions
• TM Applications - RAG

Notebooks - FineTuning

• Prompt Engineering
• PEFT - LoRA

Resources

• LangChain
• LanceDB - Vector database