Google

A common real-world application involves starting with a small, high-quality seed dataset, often created by domain experts, for a specific task like question-answering. However, this initial data is typically insufficient in both size and variety. Self-Instruct techniques can be used to address this limitation by augmenting the seed set, generating a more diverse range of fine-tuning samples. This process effectively bootstraps the LLM's performance, expanding its capabilities from a limited initial collection of data.

Bootstrapping LLMs with Self-Instruct from a Seed Dataset

Given the following scenario, propose the most effective next step for the development team and justify why this approach is suitable for their situation.

AI Assistant Development Strategy

A development team has a limited set of 200 expert-written examples for training a model to summarize legal documents. To create more training data, they use these examples as few-shot prompts for a powerful, general-purpose language model, asking it to generate thousands of new legal document summaries. Explain the primary advantage and the most significant risk of this data generation method compared to solely using the original 200 expert-written examples.

Evaluating a Data Augmentation Strategy

A team is developing a specialized chatbot to answer questions about a company's internal financial policies. They begin with a small, high-quality 'seed' dataset of 150 question-answer pairs written by their finance experts. To expand this dataset, they use the seed examples to prompt a large base model to generate 15,000 new, similar question-answer pairs. This new, larger dataset is then used to fine-tune the chatbot. Which of the following describes the most significant potential weakness of

Learn Before

Related