Infinity-Instruct

Large-scale high-quality instruction dataset for language model training. Comprises 7.4M foundational instructions (InfInstruct-F-7.4M) curated from 100M+ samples and 1.5M chat instructions (InfInstruct-G-1.5M) synthesized via instruction selection, evolution, and diagnostic filtering. 7M Core subset reaches 95.7% of full 7M performance with only 1.4M instructions. InfInstruct-LLaMA3.1-70B outperforms GPT-4-0314 by 8.6% on instruction following tasks. Updated in 2025 with instruction labeling types and reward scores.