Introduction

This post is about large language models (LLMs), specifically prompt engineering and p-tuning. LLMs are models with 1B or more parameters, and they require a large amount of training data and computational resources to be trained.

Why Use Large Language Models?

LLMs can generate text that is difficult to distinguish from text written by a human. They can also perform tasks such as translation, summarization, and question answering. LLMs have the advantage of being able to handle a wide range of tasks without retraining or fine-tuning.

Value of LLMs Over Multiple Ensembles

Ensembles of models can be less expensive than LLMs, but they require multiple skill sets to be built. In addition to the input/output from the ensemble, a data set for each individual model used in the ensemble is needed.

Prompt Engineering

Prompt engineering is the process of designing prompts to generate a specific output. There are three different strategies for prompt engineering: zero-shot prompts, few-shot prompts, and chain-of-thought prompts.

Zero-Shot Prompts

Zero-shot prompts involve prompting the model without any example of expected behavior from the model.

Few-Shot Prompts

Few-shot prompts involve prompting the model with a few examples of expected behavior. For example, a prompt that asks "What's the capital of France?" expects the model to answer "Paris".

Chain-of-Thought Prompts

Chain-of-thought prompts involve a series of prompts that build upon each other to eventually generate a specific output. These prompts can be zero-shot or few-shot.

P-Tuning

P-tuning is a method for fine-tuning an LLM on specific prompts. This allows the model to be more accurate on specific tasks and prompts. However, p-tuning requires a large amount of computational resources and can take a long time.