harshpreet931/Qwen3-4B-C-Coder-SFT-v1
harshpreet931/Qwen3-4B-C-Coder-SFT-v1 is a 4 billion parameter C-language coding specialist model, fine-tuned from Qwen3-4B-Instruct-2507. This model significantly improves instruction compliance and compile rates for C code generation, achieving a 59.0% pass@1 on CEval-priv and an 80% compile rate. It is specifically optimized for generating compilable C code that adheres to given function signatures, making it ideal for C programming tasks.
Loading preview...
Overview
harshpreet931/Qwen3-4B-C-Coder-SFT-v1 is a 4 billion parameter model built upon Qwen3-4B-Instruct-2507, specifically fine-tuned for C language code generation. This model, developed by harshpreet931, focuses on improving instruction compliance and the reliability of generated C code. It was trained using QLoRA on approximately 23,000 curated C instruction pairs, emphasizing the generation of compilable and syntactically correct C functions.
Key Capabilities and Performance
- Enhanced C Instruction Compliance: Significantly improves adherence to exact function signatures and C language conventions.
- Higher Compile Rates: Achieves an 80% compile rate on the CEval-priv benchmark and 90% on McEval-C, a substantial increase over its base model.
- Execution-Based Benchmarking: Performance is validated through execution-based benchmarks, where generated C code is compiled and run against hidden tests in a sandboxed environment.
- Specialized for C: Optimized for C17/glibc-flavored C code, making it suitable for general-purpose C programming tasks.
Limitations
- Algorithmic Ability Unchanged: The model's core problem-solving capabilities remain consistent with the base model; the fine-tuning primarily improved code format and compilability, not algorithmic prowess.
- English-only Instructions: Primarily trained on English instruction data.
- Code Review Recommended: Generated code should always be reviewed and tested, ideally with sanitizers, as it inherits potential biases and limitations from the base model.