Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AMR-Evol: Adaptive Modular Response Evolution Elicits Better Knowledge Distillation for Large Language Models in Code Generation

Ziyang Luo, Xin Li, Hongzhan Lin, Jing Ma, Lidong Bing

TL;DR

This study introduces the Adaptive Modular Response Evolution (AMR-Evol) framework, which employs a two-stage process to refine response distillation, and proves the superiority of the AMR-Evol framework over baseline response distillation methods.

Abstract

The impressive performance of proprietary LLMs like GPT4 in code generation has led to a trend to replicate these capabilities in open-source models through knowledge distillation (e.g. Code Evol-Instruct). However, these efforts often neglect the crucial aspect of response quality, relying heavily on teacher models for direct response distillation. This paradigm, especially for complex instructions, can degrade the quality of synthesized data, compromising the knowledge distillation process. To this end, our study introduces the Adaptive Modular Response Evolution (AMR-Evol) framework, which employs a two-stage process to refine response distillation. The first stage, modular decomposition, breaks down the direct response into more manageable sub-modules. The second stage, adaptive response evolution, automatically evolves the response with the related function modules. Our experiments with three popular code benchmarks (HumanEval, MBPP, and EvalPlus) attest to the superiority of the AMR-Evol framework over baseline response distillation methods. By comparing with the open-source Code LLMs trained on a similar scale of data, we observed performance enhancements: more than +3.0 points on HumanEval-Plus and +1.0 points on MBPP-Plus, which underscores the effectiveness of our framework. Our codes are available at https://github.com/ChiYeungLaw/AMR-Evol.

AMR-Evol: Adaptive Modular Response Evolution Elicits Better Knowledge Distillation for Large Language Models in Code Generation

TL;DR

This study introduces the Adaptive Modular Response Evolution (AMR-Evol) framework, which employs a two-stage process to refine response distillation, and proves the superiority of the AMR-Evol framework over baseline response distillation methods.

Abstract

The impressive performance of proprietary LLMs like GPT4 in code generation has led to a trend to replicate these capabilities in open-source models through knowledge distillation (e.g. Code Evol-Instruct). However, these efforts often neglect the crucial aspect of response quality, relying heavily on teacher models for direct response distillation. This paradigm, especially for complex instructions, can degrade the quality of synthesized data, compromising the knowledge distillation process. To this end, our study introduces the Adaptive Modular Response Evolution (AMR-Evol) framework, which employs a two-stage process to refine response distillation. The first stage, modular decomposition, breaks down the direct response into more manageable sub-modules. The second stage, adaptive response evolution, automatically evolves the response with the related function modules. Our experiments with three popular code benchmarks (HumanEval, MBPP, and EvalPlus) attest to the superiority of the AMR-Evol framework over baseline response distillation methods. By comparing with the open-source Code LLMs trained on a similar scale of data, we observed performance enhancements: more than +3.0 points on HumanEval-Plus and +1.0 points on MBPP-Plus, which underscores the effectiveness of our framework. Our codes are available at https://github.com/ChiYeungLaw/AMR-Evol.
Paper Structure (34 sections, 5 equations, 4 figures, 15 tables)

This paper contains 34 sections, 5 equations, 4 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLMs and Code Generation.
  4. Knowledge Distillation for Code Generation.
  5. Method
  6. Direct Response Distillation
  7. Adaptive Modular Response Evolution
  8. Modular Decomposition (MD).
  9. Adaptive Response Evolution (ARE).
  10. Functional Module Database.
  11. Knowledge Distillation
  12. Experiment
  13. Setup
  14. Baselines.
  15. Datasets and Benchmarks.
  16. ...and 19 more sections

Figures (4)

  • Figure 1: Direct distillation from the teacher model possibly yields low quality responses for complex tasks, thereby causing confusion within the student model.
  • Figure 2: Our Adaptive Modular Response Evolution (AMR-Evol) framework with modular decomposition and adaptive response evolution elicits better response distillation for LLMs in code generation.
  • Figure 3: Manual evaluation of the accuracy of various code response distillation methods across 120 randomly selected samples from each complexity level.
  • Figure 4: Screenshot of the interface for the human annotators to annotate whether the responses are suitable or not.