Magpie-Align
3M self-synthesized instruction pairs generated by prompting Llama 3 to produce both instructions and responses using a novel pre-query template approach. Unlike previous datasets, Magpie requires no seed data or human curation, achieving superior quality through alignment filtering.
from datasets import load_dataset
ds = load_dataset("Magpie-Align/Magpie-Pro-300K-Filtered")Preview a sample row
{
"instruction": "Explain how transformer attention mechanisms work using a library analogy.",
"response": "Imagine a library where each book (token) can query the card catalog (keys) to find related books (values). The attention mechanism...",
"model": "Meta-Llama-3-8B-Instruct"
}Fine-tune with this dataset
Estimated VRAM to fine-tune with QLoRA (4-bit base model + LoRA adapters), using conservative defaults:
New to fine-tuning? Follow the step-by-step walkthrough: Fine-Tune Your First LLM in 1 Hour
Frequently asked questions
Can I use Magpie-Align commercially?
Yes — Magpie-Align is released under Apache 2.0, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does Magpie-Align contain, and do I need all of it?
Magpie-Align contains 3M Pairs. You rarely need all of it: for style and format fine-tuning, a few hundred to a few thousand examples are enough — load a slice (e.g. split="train[:1000]") and scale up only if quality plateaus.
What is Magpie-Align best used for?
Fresh synthetic SFT data without seed data or scraping. It belongs to the Instruction / SFT section of our dataset hub, where you'll find alternatives and complementary sets.