lucas-sousa-pereira/qwen2.5-7b-java-code-refactoring

Qwen2.5-7B-Instruct Fine-Tuned for Java Code Refactoring

Overview

This model is a fine-tuned version of Qwen2.5-7B-Instruct specialized for Java code refactoring using the InstructLab framework.

The training dataset was generated using the Skills pipeline of InstructLab, with GPT-4o acting as the teacher model to generate instruction-response pairs. A total of 1,440 synthetic training instances were used for fine-tuning.

Base Model

• Qwen/Qwen2.5-7B-Instruct

Training

The dataset was generated using:

ilab data generate \
  --model "openai/gpt-4o" \
  --endpoint-url "http://localhost:8000/v1" \
  --api-key "$ILAB_API_KEY" \
  --pipeline simple

The model was fine-tuned using:

ilab model train \
  --model-path Qwen/Qwen2.5-7B-Instruct \
  --num-epochs 3 \
  --device cpu \
  --max-seq-len 2048 \
  --max-batch-len 10000 \
  --effective-batch-size 128 \
  --lora-rank 32 \
  --lora-alpha 64 \
  --lora-dropout 0.1 \
  --learning-rate 1e-4 \
  --is-padding-free false

Intended Use

This model is intended for Java code refactoring tasks. It is designed to generate refactored Java code while preserving the original program behavior.

Training Data

• Framework: InstructLab
• Pipeline: Skills
• Teacher model: GPT-4o
• Number of training instances: 1,440

License

This model is released under the Apache 2.0 license, consistent with the license of the base model.