🚀 Qwen2.5-0.5B-Code-BCP-V2

📝 Overview

This model is a fine-tuned version of Qwen2.5-0.5B-Instruct, specialized for real-time code refactoring, logging injection, and algorithmic optimization. It is designed to power VSCode extensions where low latency and local execution are critical.

Compared to the base model, BCP-V2 demonstrates an emergent understanding of time complexity (O(n) awareness) and strictly follows developer-centric instructions without unnecessary conversational filler ("Zero-Yapping").

Key Capabilities:

• Optimization: Identifying and refactoring nested loops into Hash Map lookups.
• Structured Logging: Injecting custom-formatted logs (e.g., [MONITOR] templates).
• Logic Transformation: Converting recursive functions to iterative patterns.
• IDE Ready: Optimized for GGUF format for seamless integration with Ollama or llama.cpp.

---

📊 Training Details

• Base Model: Qwen2.5-0.5B-Instruct (4-bit quantized)
• Framework: Unsloth
• Dataset: iamtarun/python_code_instructions_18k_alpaca
• Method: LoRA (Low-Rank Adaptation)
• Steps: 600 steps (~4,800 examples processed)
• Batch Size: 8 (2 per device × 4 accumulation steps)
• Scheduler: Cosine learning rate decay
• Optimizer: AdamW 8-bit

---

📈 Evaluation: V1 vs. V2 Comparison

During development, we analyzed the impact of training duration on algorithmic reasoning.

Feature	Base Model (0.5B)	BCP-V1 (150 steps)	BCP-V2 (600 steps)
Response Speed	Instant	Instant	Instant
Instruction Adherence	Medium	High	Strict
Algorithmic Reasoning	Low	Low	High (O(n) intent)
Explanations (Yapping)	High	Low	Minimal (Zero-Yapping)

Notable Improvements in V2:

• Test Case (Hash Map): While V1 failed to optimize nested loops, V2 correctly identified the need for a lookup dictionary to improve performance from $O(n^2)$ to $O(n)$.
• Test Case (Logging): V2 handles complex string interpolation (e.g., using **locals()) while maintaining strict template formatting.

---

💻 Usage

Prompt Format:

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{code_snippet}

### Response:
{code_snipped_refactored}

Running with Ollama:

1. Download the .gguf file from this repository.

2. Create a Modelfile: FROM ./qwen2.5-0.5b-instruct.Q4_K_M.gguf TEMPLATE "{{ .Prompt }}"

3. Run: ollama create bcp-v2 -f Modelfile

⚠️ Limitations

As a 0.5B parameter model, BCP-V2 is highly efficient but may occasionally produce minor syntax errors in very complex logic. It is best used for refactoring snippets of up to 50 lines and as a high-speed coding assistant.

🤝 Collaboration

This model was developed as part of a project to create an intelligent local-first VSCode extension chatbot.

Lead Fine-tuning Engineer: Alex (alex2110)