Model Summary
Qwen2.5-Coder-7B-Swift-Lm is a fine-tuned version of the Qwen2.5-Coder-7B-Instruct model, specifically optimized for high-performance Swift development.
Key Features
Modern Swift Proficiency: Deeply familiar with Swift 6 Concurrency (Actors, Sendable) and the Observation framework (@Observable).
Native SDK Knowledge: Fine-tuned to understand the nuances of SwiftUI, AppKit, and Combine.
Local Performance: Optimized for GGUF format to ensure low-latency inference on Apple Silicon (Metal) via tools like LM Studio.
Training Details
Dataset: 5,602 high-quality, curated Swift examples.
Fine-tuning Method: QLoRA (Rank=16, Alpha=32).
Target Modules: q_proj, v_proj.
Objective: Improving code generation accuracy and reducing common pitfalls like forced-unwrapping.
Training Results
Final Training Loss: 0.369
Training Epochs: 1
Loss Curve: The model showed a steady, controlled descent from an initial loss of 12.45 down to 0.36, indicating successful convergence on the Swift-specific dataset without aggressive overfitting.
Usage Instructions
LM Studio / GGUF Download the .gguf file from the Files and versions tab.
In LM Studio, load the model and enable GPU Offload to Max.
Set your system prompt to: "You are a Senior iOS Architect specializing in clean, safe, and modern Swift code."
Connecting to Cursor Start the Local Server in LM Studio (default: http://localhost:1234).
In Cursor Settings, add a custom OpenAI API with the URL http://localhost:1234/v1.
Select this model to get fine-tuned Swift suggestions directly in your project.