Google

By fine-tuning a Large Language Model on a dataset that includes instructions and their corresponding outputs across a variety of NLP problems, the model can learn to handle multiple, distinct tasks simultaneously.

Multi-Task Learning via Instruction Fine-Tuning

A startup is developing a general-purpose AI assistant using a pre-trained large language model. Their goal is to create a single model that can handle a wide variety of user requests, such as drafting emails, summarizing articles, and answering factual questions. They have a limited budget for data collection and model training. They are considering two different strategies for creating their fine-tuning dataset:

*   **Strategy A:** Create a single, unified dataset that mixes together thousands of examples of email drafting, article summarization, and question-answering, with each example framed as a specific instruction (e.g., "Summarize the following text:", "Draft an email to...").

*   **Strategy B:** Create three separate, highly specialized datasets—one exclusively for email drafting, one for summarization, and one for question-answering. They would then train three separate, specialized models.

Given the startup's goal of creating a single, versatile AI assistant with a limited budget, which strategy should they choose? Justify your recommendation by explaining the likely impact of the chosen data strategy on the model's final capabilities.

AI Assistant Fine-Tuning Strategy

A startup is developing a general-purpose AI assistant. They have two potential fine-tuning datasets to train their base language model, both of the same size.

*   **Dataset X:** Contains 500,000 high-quality examples of a single task: summarizing news articles. Each example is framed as an instruction (e.g., 'Summarize the following text: ...').
*   **Dataset Y:** Contains 500,000 examples spread across 20 different tasks, including summarization, translation, question answering, and creative writing. Each example is also framed as an instruction.

Evaluate which dataset is more suitable for creating the general-purpose AI assistant. Justify your choice by explaining how the composition of the fine-tuning data influences the model's ability to handle a wide range of user requests.

Evaluating Fine-Tuning Datasets for a General-Purpose AI

A research team fine-tunes two identical large language models. Model A is fine-tuned exclusively on 100,000 examples of text summarization, each presented as an instruction. Model B is fine-tuned on a dataset of the same total size (100,000 examples), but this dataset is a mix of summarization, translation, and question-answering tasks, all framed as instructions. When tested on a completely new task—sentiment analysis—Model B performs significantly better than Model A, which fails almost compl

Learn Before

Related