<aside> 🔥 Welcome to the Language Model Mastery Challenge, where you’ll embark on an exhilarating journey of fine-tuning your very own language model! Get ready to unleash your creativity as you dive into the vast sea of existing text and curate your own content to train your model. But that’s not all! We’ll also equip you with the skills to become a master prompter, enabling you to extract the full potential of ChatGPT-like models. And here comes the thrilling part: a virtual quest awaits as you attempt to breach the defences of a language model through the art of prompt injection, unveiling a secret password hidden within its digital labyrinth. Brace yourself for an adventure that blends learning, hacking, and sheer enjoyment as you unlock the secrets of language models. Let the challenge begin!

</aside>

A review of this challenge from a friend:

Useful things:

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

sk-proj-rvcNPn7yrnsB50KKtqvAT3BlbkFJJL2MvB17gEslrU23Ltcr

Section 1: ChatGPT and Prompt Engineering

I’m sure you’ve all heard of ChatGPT by now (play with it at chat.openai.com if not!). It’s a chatbot designed by OpenAI that can generate pretty impressive text output, just by asking it to — colourful haikus, full history essays, ASCII art, you name it. In fact, the wonderfully creative introduction to the challenge we’re doing today was written by asking ChatGPT this:

*Write a brief introductory paragraph to a challenge about language models. This challenge involves:

• Fine-tuning your own language model on both existing text and content of your choosing
• Learning how to prompt ChatGPT-like models more effectively
• Hacking into ChatGPT using prompt injection to reveal a secret password Make it sound fun, but don’t be too cringe.*

(Not sure how well it did with the last instruction, but you get the idea)

This text used to generate output from a language model is called a prompt. Despite how powerful these models are proving to be, they are totally at the mercy of the prompt — without one, they can’t predict an output!

The predictions made my large language models are hugely sensitive to the prompt context, meaning that the more information you provide, the more you can influence the way the model responds to things. Here’s a fun example.

ChatGPT does not understand basic maths

You may perhaps assume that an AI model trained at the cost of millions of pounds would be able to perform simple multiplication?

I asked ChatGPT to perform a quick calculation with some random numbers, and I was surprised at just how badly it did.

Here’s the calculation in a calculator: