Google

In the Self-Instruct process, a Large Language Model generates new instructions through in-context learning. This is achieved by providing the LLM with a selection of instructions sampled from the task pool, which serve as demonstration examples. By learning from these examples within the prompt, the model is guided to produce a new, relevant instruction.

Instruction Generation in Self-Instruct

In each iteration of the Self-Instruct process, a small subset of instructions is selected from the task pool to be used as prompts for generating new instructions. To maintain diversity in the generated tasks, this selection can include a mix of both the initial, human-written seed instructions and the instructions previously generated by the Large Language Model.

Instruction Sampling for Diversity in Self-Instruct

An example of a prompt template used to generate a new instruction in the Self-Instruct process is: "You are provided several different instructions for performing some tasks. Please generate an instruction based on these. Task 1: $${\mathrm{instruction}1}$$ Task 2: $${\mathrm{instruction}2}$$ Task 3: $${\mathrm{instruction}3}$$ Task 4: $${\mathrm{instruction}4}$$ New Task:". This template directs the Large Language Model to synthesize a new task instruction by learning from the provided examples in-context.

Example of a Prompt Template for Instruction Generation in Self-Instruct

Following the creation of a new instruction within the Self-Instruct framework, a Large Language Model is prompted to generate a complete sample. This process involves using the newly generated instruction to guide the model in creating the necessary input fields and generating the corresponding output, thereby completing the data instance.

Sample Generation in Self-Instruct

A research team is building a dataset to train a model for generating marketing slogans. They start with a small seed set of 10 human-written instructions. To expand this set, they provide a large language model with the following prompt:

'Here are some examples of instructions:
1. Create a slogan for a new electric car.
2. Write a tagline for a vegan restaurant.
3. Invent a short, memorable phrase for a new fitness app.

Based on these, please generate a new and different instruction for a similar task.'

The model then outputs: 'Generate a catchy tagline for a new brand of eco-friendly coffee.'

Explain the underlying mechanism that enables the model to produce this new, relevant instruction based on the prompt it received.

Automated Task Creation for a Marketing Dataset

A team wants a large language model to generate a new, high-quality instruction for a customer service chatbot. They provide the model with the following prompt: 'Please generate a new instruction for a chatbot.' The model's output is generic and not very useful. Based on the principles of generating new tasks from existing ones, analyze why this approach is likely to fail and describe what key component is missing from the prompt.

Analyzing a Flawed Prompt for Instruction Generation

An AI development team provides a large language model with a prompt containing several existing task instructions, such as 'Translate this sentence into French' and 'Write a poem about the ocean.' The prompt then asks the model to generate a new, distinct instruction based on the examples provided. What is the primary function of including the existing instructions in the prompt?

Learn Before

Related