LM-Mixup enhances LLM instruction tuning by distilling low-quality data

Researchers have introduced LM-Mixup, a novel method for augmenting text data to improve instruction tuning in Large Language Models (LLMs). This technique addresses the scarcity of high-quality data by distilling abundant, low-quality or redundant data into coherent instruction-output pairs. The LM-Mixup process involves supervised fine-tuning on a distilled dataset called MIXTURE, followed by reinforcement learning using quality, semantic alignment, and format compliance rewards via Group Relative Policy Optimization (GRPO). Experiments show that LLMs fine-tuned on data augmented with LM-Mixup outperform those trained on the full dataset and rival state-of-the-art data selection methods, demonstrating the value of distilled low-quality data. AI

IMPACT Enhances LLM efficiency and performance by enabling better utilization of low-quality data for instruction tuning.

RANK_REASON The cluster contains a research paper detailing a new method for data augmentation in LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Group Relative Policy Optimization
• Hugging Face
• Large Language Models
• LM-Mixup
• MIXTURE
• Zhijie Deng

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →