Qwen3-4B-Coder

A fine-tuned Qwen3-4B model specialized for Python code generation, trained by an elementary school student on an RTX 4060 Laptop GPU (8 GB VRAM).

Qwen3-4Bをベースに、Pythonコード生成に特化してファインチューニングしたモデルです。小学生がRTX 4060 Laptop GPU (VRAM 8GB) で学習しました。

Benchmark Results

MBPP-sanitized (Practical Python Tasks)

Model	MBPP pass@1	Condition
Qwen3-4B-Coder (this model)	69.3% (178/257)	Q4_K_M, temperature=0.0
Qwen3-4B (official)	62.0%	FP16, EvalPlus

+7.3 points improvement on practical coding tasks.

HumanEval (Algorithmic Tasks)

Model	HumanEval pass@1	Condition
Qwen3-4B-Coder (this model)	47.6% (78/164)	Q4_K_M, temperature=0.0
Qwen3-4B (official)	65.6%	FP16, EvalPlus

Inference Speed

Benchmark	Qwen3-4B-Coder	Qwen3-4B (Q4_K_M)	Speed Ratio
HumanEval (164 tasks)	793s	3623s	4.6x faster
MBPP (257 tasks)	1274s	-	-

Syntax error rate on HumanEval: 0% (164/164)

Key Findings

This model demonstrates that SFT for code-only output has two major benefits:

1. Practical code generation ability improved (MBPP +7.3 points)
2. Inference speed improved 4.6x by eliminating think blocks and explanations

Training Details

Parameter	Value
Base Model	Qwen/Qwen3-4B
Method	SFT with LoRA (r=16, alpha=32)
Dataset	PersonalAILab/AFM-CodeAgent-SFT-Dataset
Training Samples	8,869 (filtered to 512 tokens)
Epochs	3
Final Loss	0.72
MAX_SEQ	512
GPU	NVIDIA RTX 4060 Laptop (8 GB VRAM)
Training Time	~5.5 hours
Quantization	Q4_K_M (~2.4 GB)

Features

• Code-only output without extra explanations
• 4.6x faster inference than base model
• Supports English and Japanese prompts
• Optimized for agent pipelines
• Syntax error rate 0% on HumanEval

AI Code Agent (CLI Tool)

An interactive CLI tool that uses this model to generate, execute, and auto-fix Python code.

git clone https://github.com/jiexiang018-tech/ai-python-agent.git
cd ai-python-agent
pip install -r requirements.txt
python setup.py
python agent.py