nurl-lang/Qwen3-4B-NURL
Qwen3-4B-NURL is a 4 billion parameter language model developed by nurl-lang, fine-tuned from the Qwen3-4B base model. It is specifically optimized for generating idiomatic NURL code, a compiled systems programming language with terse prefix notation and compile-time ownership. This model excels at NURL function completion, documentation Q&A, and generating NURL source code, making it ideal for developers working with the NURL language.
Loading preview...
Model Overview
nurl-lang/Qwen3-4B-NURL is a specialized 4 billion parameter language model derived from the Qwen3-4B base. Its core purpose is to generate correct and idiomatic code in NURL, a systems programming language characterized by its terse prefix notation and compile-time ownership. Unlike the base Qwen3-4B, which lacks NURL knowledge, this fine-tuned version produces syntactically and semantically accurate NURL code examples.
Key Capabilities
- NURL Code Generation: Produces idiomatic NURL functions and code snippets.
- Contextual NURL Understanding: Capable of function completion, answering questions based on NURL documentation, and generating code from specifications.
- Efficient Training: Fine-tuned using
nurllama, a pure-NURL LLM engine, on a single RTX 4090 GPU. - LoRA Adapters: Utilizes LoRA (rank 16, α 32) applied to all 36 layers (q/k/v/o + gate/up/down projections) for efficient adaptation.
- High-Quality Training Data: Trained on a 7.5 million-token corpus derived exclusively from the NURL repository, including chat samples, function completions, documentation Q&A, and raw source code, all from ground truth sources.
Usage and Deployment
This model can be used directly with the Hugging Face transformers library, including a pre-configured chat template. Alternatively, users can download the compact 132 MB LoRA adapters and merge them locally with the base Qwen3-4B GGUF model using nurllama, offering flexibility for deployment and integration into NURL-native workflows.