Code Standard Agent — Fine-Tuned Code-Quality Model

This project builds a fine-tuned DeepSeek Coder model capable of taking raw “messy” Python code and generating a clean, standardized, and fully equivalent Python version. The goal is to help developers maintain clean, Pythonic, and professional-quality code while preserving functional behavior.

---

Project Overview

This project fine-tunes a DeepSeek Coder model using PyTorch, HuggingFace Transformers, and LoRA/QLoRA.
The model learns to:

• Detect common code anti-patterns
• Recognize messy or poorly structured Python code
• Produce clean, PEP-8–aligned replacements
• Maintain identical functional behavior

Training uses a 2,600-example paired “messy → clean” dataset synthesized with Claude Haiku and validated to guarantee functional equivalence.

---

Dataset

Dataset Size:

2,600 paired examples (messy_code → clean_code)

How the dataset was generated

I built a custom synthetic dataset generator using Claude models:

• Claude Haiku produces intentionally messy Python code:
  • Bad variable names
  • Dead code
  • Misleading comments
  • Redundant logic
  • Indentation issues
  • Unnecessary variables
• Claude Haiku then rewrites the messy code using strict rules:
  • Must preserve exact behavior
  • Must not modify algorithms, return values, or logic
  • Only structural and stylistic improvements allowed

Dataset Generation Script

The generation.py creates synthetic pairs using Claude Haiku with behavior-preservation constraints.

Key features:

• Messy code generator
• Clean code generator
• JSON + JSONL writing
• Automatic appending to existing dataset

---

Model Finetuning

The training uses:

• DeepSeek Coder 1.3B Base
• PyTorch
• HuggingFace Transformers
• LoRA via peft (efficient parameter tuning)
• QLoRA (4-bit quantization for GPU efficiency)