Data Quality and Diversity
You'll learn why the quality and variety of your instruction dataset directly determine how capable and robust your instruction-tuned model becomes.
The Foundation of Instruction Tuning Success
After you've built instruction datasets and started supervised fine-tuning (lessons 1751, 1753), a critical question emerges: not all instruction data is created equal. Two models trained with the same architecture but different instruction datasets can perform drastically differently.
Data quality refers to how well each instruction-response pair demonstrates the desired behavior. High-quality examples have clear instructions, accurate responses, proper formatting, and appropriate reasoning patterns. Low-quality data includes ambiguous instructions, factually incorrect responses, inconsistent formatting, or responses that don't actually follow the instruction.
Data diversity means covering a broad range of tasks, domains, instruction phrasings, and complexity levels. A dataset with 10,000 examples of simple question-answering teaches your model to be a narrow specialist. A dataset with 10,000 examples spanning mathematics, coding, creative writing, analysis, and reasoning produces a more capable generalist.
Why Both Matter
Think of it like learning to cook: if you only practice making scrambled eggs (low diversity), you'll master eggs but fail at pasta. If you practice many dishes but always use spoiled ingredients (low quality), you'll learn bad techniques. You need both variety and excellence.
Quality affects correctness and reliability. Diversity affects generalization and versatility. A model trained on high-quality but narrow data won't handle new task types. A model trained on diverse but low-quality data will confidently give wrong answers across many domains.
Key Takeaway: The quality of individual instruction-response pairs determines how well your model learns each task, while diversity across tasks determines how broadly useful your instruction-tuned model becomes—both are essential for building capable AI assistants.