Google

Supervised Fine-Tuning (SFT) datasets are composed of diverse instruction-response pairs to teach models a variety of tasks. Examples of such pairs include:
*   **Summarization:** Given an instruction to summarize an article and the article's text, the model should produce a concise summary.
*   **Information Extraction:** Given an instruction to extract financial figures from a report and the report's text, the model should output the specific figures like revenue and profit margin.
*   **Classification:** Given an instruction to classify an email and the email's text, the model should output the correct category, such as "Spam".
*   **Problem-Solving:** Given an instruction to provide a solution to a technical issue and a description of the issue, the model should generate helpful troubleshooting steps.

Examples of Instruction-Following Tasks in SFT Datasets

Match each example of an instruction-response pair to the type of instruction-following task it represents.

Which type of instruction-following task would be most appropriate for creating the fine-tuning data to meet the company's goal described in the case study? Justify your choice by explaining how the instruction-response pairs for this task would be structured.

Selecting a Fine-Tuning Task for a Business Need

You are creating a dataset to fine-tune a language model. Your goal is to teach the model how to perform an 'Information Extraction' task. Construct a single, complete data sample for this task. Your sample must include: 1) a clear instruction for the model, 2) a short paragraph of input text (2-4 sentences), and 3) the ideal, extracted output the model should generate.

Learn Before

Related