bjivanovich/swift-1.5-27b-coding-GGUF

swift-1.5-27b-coding-GGUF

Official GGUF quantized weights for [swift-1.5-27b-coding](https://huggingface.co/ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP), a specialized 27B parameter coding model fine-tuned using DoRA on bjivanovich/code-py-rust-cpp-50k (covering Python, Rust, C++, and multi-step reasoning).

All quantizations were generated using an importance matrix (imatrix) calibrated directly on domain-specific programming samples.

Quantization Details

Training & Specialization

• —Base Architecture: 27B Parameters
• —Technique: DoRA (Weight-Decomposed Low-Rank Adaptation)
• —Dataset: bjivanovich/code-py-rust-cpp-50k (50,000 samples)
• —Python: 59.0%
• —C++: 20.4%
• —Rust: 15.5%
• —Reasoning: 5.0%

Prompt Template (ChatML)

<|im_start|>user
Write a thread-safe generic queue in Rust using Arc and Mutex with unit tests.<|im_end|>
<|im_start|>assistant