Loading Now

Summary of Xft: Unlocking the Power Of Code Instruction Tuning by Simply Merging Upcycled Mixture-of-experts, By Yifeng Ding et al.


XFT: Unlocking the Power of Code Instruction Tuning by Simply Merging Upcycled Mixture-of-Experts

by Yifeng Ding, Jiawei Liu, Yuxiang Wei, Terry Yue Zhuo, Lingming Zhang

First submitted to arxiv on: 23 Apr 2024

Categories

  • Main: Computation and Language (cs.CL)
  • Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Software Engineering (cs.SE)

     Abstract of paper      PDF of paper


GrooveSquid.com Paper Summaries

GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!

Summary difficulty Written by Summary
High Paper authors High Difficulty Summary
Read the original abstract here
Medium GrooveSquid.com (original content) Medium Difficulty Summary
The paper introduces XFT, a training scheme that combines upcycled Mixture-of-Experts (MoE) with Large Language Models (LLMs). While vanilla sparse upcycling fails to improve instruction tuning, XFT incorporates a shared expert mechanism and routing weight normalization strategy, boosting performance. The approach involves fine-tuning an upcycled MoE model and then merging it into a dense model, achieving state-of-the-art results on HumanEval and HumanEval+ with a 1.3B model. XFT also improves supervised fine-tuning (SFT) by 13% on HumanEval+, demonstrating its generalizability. The approach is orthogonal to existing techniques like Evol-Instruct and OSS-Instruct, opening new avenues for improving code instruction tuning.
Low GrooveSquid.com (original content) Low Difficulty Summary
The researchers created a new way to train language models using a technique called XFT. This method combines two things: upcycled Mixture-of-Experts (MoE) and Large Language Models (LLMs). The result is a model that can learn more quickly and accurately than previous models. They tested this approach on several tasks and found it improved the results by 13%. What’s exciting about XFT is that it doesn’t replace other techniques, but rather works alongside them to make language models even better.

Keywords

» Artificial intelligence  » Boosting  » Fine tuning  » Instruction tuning  » Mixture of experts  » Supervised