한국어EN
Training Data & Post-TrainingXu et al.

Magpie

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

Magpie — 4 million instructions generated from nothing but an empty chat template

Why it matters

Distillation became so cheap that it no longer even needs API calls

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2406.08464

Added in Week 35, 2026 · See that week

On the map — ⑫ Training Data line

  1. Previous stopRunning out of data?
  2. This recordMagpie
  3. Next stopCan AI scaling continue through 2030?
See it on the map

Same week

SB Intuitions releases Sarashina1 series 7B/13B/65B Japanese LLMsOfficialMaxime Labonne publishes AbliterationOfficialOpenVLAPaperReconciling Kaplan and Chinchilla Scaling LawsPaper

Nearby on the same axis

Nemotron-4 340BPaperWill we run out of data?OfficialFineWeb paper and FineWeb-Edu, 1.3 trillion tokensPaperSimPOPaper