gcelikmasat-work/Qwen3_4B_BPMN_IT
Qwen3-4B-InstruBPM by gcelikmasat-work is a 4 billion parameter, instruction-tuned language model based on Qwen3-4B-Instruct-2507, specifically designed to convert natural language business process descriptions into BPMN models rendered in Graphviz DOT. It achieves near-perfect structural fidelity (R-GED Accuracy "," 99.4%) and outperforms larger proprietary and open-weight models on key metrics for this specialized task, making it highly effective for accelerating early-stage BPMN model generation. The model supports a subset of BPMN elements including tasks, events, sequence flows, and AND/XOR gateways, with a context length of 32768 tokens.
Loading preview...
Qwen3-4B-InstruBPM: Natural Language to BPMN Model Generation
Qwen3-4B-InstruBPM, developed by gcelikmasat-work, is a specialized 4 billion parameter instruction-tuned language model. It is a LoRA adaptation of Qwen/Qwen3-4B-Instruct-2507, uniquely engineered to transform natural language descriptions of business processes directly into BPMN models, outputting them in Graphviz DOT format.
Key Capabilities
- High Fidelity BPMN Generation: Achieves exceptional structural accuracy (R-GED Accuracy "," 99.4%) on a stratified 180-instance benchmark, demonstrating robust conversion from text to BPMN.
- Outperforms Larger Models: Despite its compact 4B parameter size, it matches or surpasses untuned open-weight baselines (e.g., Qwen2.5 7/14B, Qwen3 30B) and strong proprietary systems (e.g., GPT-5.1, Claude 4.5, Gemini 2.5 Pro/Flash) on BLEU, ROUGE-L, and METEOR scores for BPMN generation.
- Efficient: Delivers high performance at roughly half the parameter count of prior tuned models for this task.
- Supported BPMN Elements: Generates start/end events, tasks, sequence flows, and AND/XOR gateways.
Good For
- Business Process Modelers and Analysts: Ideal as an assistant for generating first-draft BPMN models from textual descriptions, significantly accelerating the early stages of process modeling.
- Automating Initial BPMN Design: Useful for quickly visualizing process flows described in natural language, reducing manual effort.
- Research and Development: Provides a strong baseline for further research into text-to-BPMN generation and domain-specific LLM fine-tuning.
Limitations
Currently, the model focuses on the control-flow slice of BPMN and does not generate pools, lanes, message flows, data objects, sub-processes, or annotations. It is trained on English only, and generalization to diverse enterprise documentation may require further validation. Human review of generated models is recommended, especially for complex gateway logic and activity labels.