A LoRA fine-tuned adapter on top of Qwen/Qwen2.5-3B-Instruct that converts natural language system descriptions into our formal structural architectural design language, SADL.
SADL serves as an intermediate design language to represent systems to be modeled in AADL (Architecture Analysis & Design Language) and SysML (Systems Modeling Language).
This adapter was developed as part of the SysGenAI project, which pairs this model with a Java-based SDD generator to produce complete architecture diagrams from plain-text input.
Model Details
Model Description
Developed by: chambersc2017
Model type: LoRA adapter (PEFT) on Qwen2.5-3B-Instruct
This model is intended to be used as part of the SysGenAI pipeline. Given a natural language description of a system, the model generates corresponding SADL translation, which can then be rendered as into architecture diagrams (AADL/SysML).
Demonstration:
English natural language: "Inside the coffeemaker, two components reside: a cooking unit and a storage unit."
SADL: "COFFEEMAKER consists internal_components : COOKING_UNIT and STORAGE_UNIT."
English natural language: "The power unit houses five internal components: a cord, a power switch, a junction box, heating wires, and motor wires."
English natural language: "A filter holder and a filter insert are recognized as the internal components of the HEPA assembly."
SADL: "HEPA_ASSEMBLY consists internal_components : FILTER_HOLDER and FILTER_INSERT."
Example use cases:
Generating AADL component definitions from a system narrative
Producing SysML block diagrams from requirements text
Rapid prototyping of system architecture models without manual modeling effort
Downstream Use
The adapter is designed to slot into the SysGenAI project pipeline:
User provides a plain-English system description as input
This model generates formal SADL translation
Output is passed to SDD_Generator.java for diagram rendering
Final diagrams are written to src/gen/aadl/ or src/gen/sysml/
Out-of-Scope Use
Not intended for general-purpose text generation or chat
Not validated for safety-critical or production-grade architecture modeling without human review
Not suitable for non-English input descriptions
Not designed for large or highly complex system architectures
How to Get Started with the Model
Install dependencies:
pip install transformers peft torch
Load the adapter:
python
1from peft import PeftModel
2from transformers import AutoModelForCausalLM, AutoTokenizer
34base_model_id ="Qwen/Qwen2.5-3B-Instruct"5adapter_id ="chambersc2017/sadl_qwen2.5_3b"67tokenizer = AutoTokenizer.from_pretrained(base_model_id)8base_model = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype="auto")9model = PeftModel.from_pretrained(base_model, adapter_id)1011prompt ="Describe a simple sensor system with a data processor and output interface."12inputs = tokenizer(prompt, return_tensors="pt")13outputs = model.generate(**inputs, max_new_tokens=512)14print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Note: This adapter is used within the SysGenAI project pipeline. For full diagram generation, run it alongside SDD_Generator.java as described in the project README.
Training Details
Training Data
The model was fine-tuned on a 50,000-pair synthetically generated dataset created specifically for the SysGenAI project. The dataset consists of paired examples of:
Natural language descriptions of the system
Corresponding SADL representations
The synthetic data was designed to cover a range of system architectures, including component hierarchies, data flows, and interface definitions.
Training Procedure
Method: LoRA (Low-Rank Adaptation) via PEFT
Base model: Qwen2.5-3B-Instruct
Hardware: NVIDIA GeForce RTX 5090 (32GB VRAM)
Driver Version: 580.126.09
CUDA Version: 13.0
Framework: PEFT 0.18.1
Training Hyperparameters
Parameter
Value
Training regime
fp16/bf16 mixed precision
LoRA rank (r)
32
LoRA alpha
64
Learning rate
2e-5
Epochs
3
Target Modules
q_proj, k_proj, v_proj, o_proj
Batch Size
2
Evaluation
Testing Data
Evaluation was performed on a held-out split of the synthetic dataset, covering system descriptions not seen during training.
Metrics
BLEU score
BLEU (Bilingual Evaluation Understudy) is a precision-focused text generation metric that evaluates machine-generated translations by measuring n-gram overlap against one or more human reference translations.
Accuracy
Accuracy is the proportion of predictions that exactly match the reference.
Parsability
Parsability refers to the degree to which a machine translation adheres to the syntactic and structural constraints defined by the SADL grammar, ensuring that the output can be deterministically parsed and interpreted by systems that implement SADL.
Results
Evaluated on 300 handwritten natural language descriptions:
Metric
Value (%)
BLEU
96.20
Accuracy
84.00
Parsability
91.67
Bias, Risks, and Limitations
Known Limitations
English only: The model has only been tested on English-language system descriptions. Non-English input is not supported and may produce unpredictable output.
Simple to moderate systems only: The model performs best on smaller, well-scoped system descriptions. Complex, highly nested, or large-scale architectures may result in incomplete or malformed output.
May hallucinate invalid syntax: As with all language models, this adapter can generate SADL translations that appears structurally plausible but contains syntax errors or semantically incorrect constructs. All output should be reviewed by a domain expert before use.
Recommendations
Always validate generated SADL output using a domain-appropriate parser or linter before integrating into a formal design
Use this model as a drafting aid, not a replacement for expert architecture modeling
Keep input descriptions concise and focused on a single system or subsystem for best results
Citation
If you use this model in academic work, please cite the base model and this adapter:
bibtex
1@misc{sadl_qwen2025,
2 author = {chambersc2017},
3 title = {SysGenAI: Qwen2.5-3B LoRA Adapter for NL-to-SADL Translation},
4 year = {2026},
5 publisher = {Hugging Face},
6 howpublished = {\url{https://huggingface.co/chambersc2017/sadl_qwen2.5_3b}}
7}
Model Card Contact
For questions or issues, open a discussion on this repository.
