codemate-qwen-1.5B

CodeMate-Qwen

Model Details

Model Description

CodeMate-Qwen is a coding-focused language model fine-tuned from Qwen2.5-Coder-1.5B using Low-Rank Adaptation (LoRA). The model is designed to assist developers with code generation, debugging, code explanation, refactoring, and software engineering tasks.

The project was created to explore parameter-efficient fine-tuning techniques and build a lightweight coding assistant capable of supporting real-world development workflows.

Developed by

Michael Moses

Funded by

Self-funded personal research project.

Shared by

Michael Moses

Model Type

Causal Language Model (LLM) for Code Generation and Software Engineering Assistance.

Language(s)

• English

• Programming Languages:

  • Python
  • JavaScript
  • TypeScript
  • HTML
  • CSS
  • SQL
  • General programming concepts

License

Apache 2.0 (subject to the licensing terms of the base Qwen model).

Finetuned From

Qwen/Qwen2.5-Coder-1.5B

---

Model Sources

Repository

GitHub: https://github.com/micymike

Hugging Face

https://huggingface.co/micymike

Demo

Coming Soon

---

Uses

Direct Use

This model is intended for:

• Code generation
• Debugging assistance
• Programming education
• Code explanation
• Refactoring recommendations
• Developer productivity workflows
• AI-assisted software development

Downstream Use

Potential downstream applications include:

• Coding copilots
• Educational coding assistants
• Automated code review systems
• Software engineering support tools
• Programming tutors

Out-of-Scope Use

This model is not intended for:

• Legal advice
• Medical advice
• Financial decision-making
• Safety-critical systems
• Autonomous code deployment without human review

Generated code should always be reviewed and tested before production use.

---

Bias, Risks, and Limitations

Like all large language models, CodeMate-Qwen may:

• Generate incorrect code
• Produce insecure implementations
• Hallucinate APIs or libraries
• Miss edge cases
• Reflect biases present in training data

Users should validate all generated outputs before deployment.

---

Recommendations

The model performs best when:

• Prompts are clear and specific
• Sufficient context is provided
• Outputs are reviewed by a developer

The model should be considered an assistant rather than a replacement for software engineering expertise.

---

How to Get Started

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "micymike/codemate-qwen-merged"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto"
)

prompt = "Write a Python function that checks if a number is prime."

inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=256
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Training Details

Training Data

The training dataset consisted of instruction-response pairs focused on software engineering and programming-related tasks.

Examples included:

• Bug fixing
• Code generation
• Code explanation
• Refactoring
• Programming Q&A
• Developer workflow assistance

Training Procedure

The model was fine-tuned using LoRA (Low-Rank Adaptation), allowing efficient adaptation of the base model while training only a small subset of parameters.

Training Regime

• Base Model: Qwen2.5-Coder-1.5B
• Fine-Tuning Method: LoRA
• Framework: Hugging Face Transformers
• PEFT Library: PEFT
• Backend: PyTorch

---

Evaluation

Testing Data

Evaluation was performed using programming-related prompts covering:

• Python debugging
• Code generation
• Code explanation
• Refactoring tasks

Metrics

Evaluation focused primarily on qualitative assessment:

• Instruction-following capability
• Code correctness
• Response quality
• Programming relevance

Results

The model demonstrated improved performance on coding-focused tasks compared to the untuned base model and showed stronger alignment with software engineering workflows.

---

Environmental Impact

Hardware Type

NVIDIA GPU

Cloud Provider

Google Colab

Compute Region

Not specified

Carbon Emitted

Not measured

---

Technical Specifications

Model Architecture

Transformer-based autoregressive language model.

Base Architecture

Qwen2.5-Coder-1.5B

Objective

Next-token prediction optimized for coding and software engineering tasks.

---

Compute Infrastructure

Hardware

Google Colab GPU Environment

Software

• Python
• PyTorch
• Transformers
• PEFT
• Hugging Face Hub

---

Citation

@misc{moses2026codemateqwen,
  author = {Michael Moses},
  title = {CodeMate-Qwen: A LoRA Fine-Tuned Coding Assistant Based on Qwen2.5-Coder-1.5B},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/micymike}
}

---

Model Card Authors

Michael Moses

---

Contact

GitHub: https://github.com/micymike

Email: [email protected]

---

Future Work

Planned improvements include:

• Larger instruction datasets
• Quantized deployments
• Benchmark evaluation on HumanEval and MBPP
• Additional programming language support
• Interactive web demo
• Advanced code review capabilities