magpie

Why we included this project

Building instruction-tuning data for your own model usually means writing seed questions and prompting an LLM to expand them into conversations. Magpie skips that manual step: it picks up the pre-query template an aligned model already uses and lets that model generate both the user question and its answer. That removes a chunk of hand curation and scales to millions of pairs, which matters when you're fine-tuning or aligning a chat model. The repository includes runnable generation scripts for Llama-3.3, Qwen2.5, Phi-3, and Gemma-2, plus released datasets you can either use directly or study as format examples before running your own. If you need supervised fine-tuning or preference-optimization data and want a less prompt-heavy pipeline, this is a solid, reproducible starting point backed by published artifacts.