Model Evaluation: LeetCode Fine-Tuning Comparison

This document presents a comparative analysis of three model versions evaluated on classic LeetCode problems. The goal was to observe the transition from a base "completion" model to a specialized coding assistant using Supervised Fine-Tuning (SFT).

🔗 Quick Links

• Model Repository: Hugging Face
• Base Model: unsloth/Qwen2.5-Coder-0.5B-bnb-4bit

🛠 Training Methodology

The models were trained using Supervised Fine-Tuning (SFT). This technique involves training the base model on a high-quality dataset of instruction-output pairs (Problem Statement -> Python/C++ Solution), allowing the model to learn both the "assistant" conversational format and specific algorithmic logic.

---

Performance Summary

Feature	Base Model	QLoRA Fine-Tune	Full Fine-Tune
Output Style	Raw code / Completion	Conversational + Code	Concise + Structured
Instruction Following	Low (Loops/Repeats)	High (Follows prompts)	Very High
Language Choice	C++ (Default)	Python	Python
Stability	Prone to repetition loops	Moderate (Some suffix noise)	High stability

---

Detailed Results

1. Two Sum

Problem: Find indices of two numbers that add up to a target.

• Base Model: Provided a standard C++ nested loop solution. Functional but lacks explanation.
• QLoRA Model: Switched to Python. Implemented an optimized $O(n)$ hash map approach.
• Full Fine-Tune: Provided a clean, professional Python implementation with focused comments on time complexity.

2. Longest Substring Without Repeating Characters

Problem: Find the length of the longest substring with unique characters.

• Base Model: Failed. Entered a repetition loop, repeating the problem description and Chinese characters indefinitely.
• QLoRA Model: Success. Implemented a sliding window approach.
• Full Fine-Tune: Success. Implemented an elegant sliding window solution with clear variable naming (longest_streak).

3. Median of Two Sorted Arrays

Problem: Find the median in $O(\log (m+n))$ time.

• Base Model: Failed. Repeated the "You are a competitive programmer" persona prompt without generating code.
• QLoRA Model: Used a sorted() approach. While functional, it operates at $O(N \log N)$.
• Full Fine-Tune: Attempted a more complex approach using heapq (min-heaps), showing a better grasp of advanced data structures.

---

Observations & Analysis

The "Repetition" Trap

The Base Model struggled significantly with the "assistant" format. Because it is a base-completion model, it often got stuck in loops. SFT successfully bridged this gap, teaching the model to recognize when to stop generating code.

QLoRA vs. Full Fine-Tuning

• QLoRA: Showed a dramatic improvement in "personality." However, it exhibited some "suffix noise" (repeating _equalTo or (egt).
• Full Fine-Tuning: Produced the most stable results. The outputs were concise, free of repetitive suffixes, and the logic was generally more robust across all 5 samples.

---

Acknowledgments & Licensing

• Base Model: This project is improved using Qwen2.5-Coder-0.5B by Alibaba Cloud.
• Dataset: Fine-tuned on the newfacade/LeetCodeDataset.
• License: The original code and fine-tuning scripts in this repository are licensed under the MIT License. The model weights are subject to the Qwen License Agreement.

---