Instruction Tuning

Domain: LLM Concepts

🎯 Motivation

Fine-tuning specifically on instruction-following data - ‘translate this’, ‘summarize this’, ‘write code for this’ - teaches the model to follow directions.

📋 Concrete Example

Dataset: [(‘Translate to French: Hello’, Bonjour), (‘Summarize: [long text]’, ‘[short summary]’)]. The model learns to follow different instruction types.

🔗 Analogy

A new employee going through orientation - they learn specific procedures: ‘when the phone rings, do this’, ‘when a customer arrives, do that’. Instruction tuning trains the model on specific tasks.

Related Concepts

• Fine-Tuning
• Supervised Fine-Tuning
• Instruction Data