RichWoollcott/DCL-Qwen3-4B-Instruct-2507
RichWoollcott/DCL-Qwen3-4B-Instruct-2507 is a 4 billion parameter Qwen3-based instruction-tuned model developed by Richard Woollcott. This model is specifically fine-tuned to author and repair DCL (Declarative Capability Language), a closed-vocabulary language for modeling business-system capabilities. It excels at generating DCL source code that compiles cleanly on the first attempt, achieving a 7/9 success rate in protocol authoring compared to 0/9 for the stock base model. Designed to run efficiently on developer laptops, it is ideal for integrating DCL authoring and repair into development workflows.
Loading preview...
Model Overview
RichWoollcott/DCL-Qwen3-4B-Instruct-2507 is a 4 billion parameter instruction-tuned model based on Qwen3, developed by Richard Woollcott. Its primary purpose is to facilitate the authoring and repair of DCL (Declarative Capability Language), a specialized, closed-vocabulary language for defining business system capabilities. The model is optimized for efficiency, with a GGUF build of approximately 2.5 GB, making it suitable for local execution on standard developer laptops.
Key Capabilities and Performance
This fine-tuned model demonstrates significant improvement over its base counterpart in DCL generation. On a compiler-graded exam, it achieved a 7/9 success rate for first-attempt clean compiles in 'protocol authoring' (where vocabulary reference is provided), a substantial increase from the base model's 0/9. It also maintains the base model's perfect 3/3 repair skill. This performance is crucial for developers working with DCL, as it drastically reduces the need for manual corrections.
Usage and Limitations
To achieve optimal performance, the model requires a specific usage protocol:
- System Prompt: "Output ONLY the DCL source. No prose, no explanation, no markdown fences."
- User Prompt: Must include the DCL vocabulary reference alongside the feature brief, as the model (like others tested) cannot author DCL from memory without it.
- Repair: Provide the broken
.dclcode and compiler diagnostics for correction.
It's important to note that the model scores 0/9 in zero-shot authoring (without the vocabulary sheet). While highly effective for DCL, it is trained on English feature briefs across synthetic business domains and DCL v1.0 only. Users should always run the model's output through the DCL compiler, as the overall design assumes a compiler in the loop for validation.