gemma-2-9b-it-BPMN

Gemma2-9B-BPMG-IT

Gemma2-9B-BPMG-IT is an instruction-tuned language model that converts natural-language business process descriptions into BPMN models rendered in Graphviz DOT. It is a LoRA adaptation of google/gemma-2-9b-it, trained on a cleaned subset of the MaD dataset for the paper:

Generating Business Process Models with Open Source Large Language Models using Instruction Tuning. Gökberk Çelikmasat, Atay Özgövde, Fatma Başak Aydemir. International Conference on Product-Focused Software Process Improvement (PROFES 2025), Springer LNCS, pp. 269–284. DOI: 10.1007/978-3-032-12089-2_17

This model is the subject of our PROFES 2025 conference paper, in which we introduced the instruction-tuning approach and evaluated it against open-weight and proprietary baselines on textual and structural metrics. It is also the BPMG-IT baseline in our subsequent journal paper:

InstruBPM: Instruction-Tuning Open-Weight Language Models for BPMN Model Generation. Çelikmasat, Özgövde, Aydemir. Software and Systems Modeling, under review, 2026. arXiv: 2512.12063

For new projects we recommend the successor model, gcelikmasat-work/Qwen3_4B_BPMN_IT, which matches this model's accuracy with roughly half the parameter count (4B vs. 9B) and ships with quantized and merge-scale variants for deployment trade-offs.

Results

Evaluated on the 180-instance stratified benchmark used in the InstruBPM journal paper (Table 2), this model attains:

These scores are very close to those of the newer 4B Qwen3 successor (which reaches 83.06 / 94.43 / 92.82 / 99.44 on the same benchmark), while requiring more than twice the memory and compute at inference time.

Intended use

Generate first-draft BPMN models from textual process descriptions to accelerate early-stage modeling. Intended as an assistant for business process modelers and analysts; human review remains recommended, particularly for gateway logic and activity labels.

Supported BPMN subset

The model generates BPMN process fragments in DOT notation covering: start events, end events, tasks (activities), sequence flows, and AND/XOR gateways (splits and joins). It does not generate pools, lanes, message flows, data objects, intermediate/boundary events, sub-processes, or annotations.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "gcelikmasat-work/gemma-2-9b-it-BPMN"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

instruction = (
    "You are an expert in BPMN modeling and DOT language. Your task is to "
    "convert detailed textual descriptions of business processes into accurate "
    "BPMN model codes written in DOT language. Label all nodes with their "
    "activity names. Represent all connections between nodes without labeling "
    "the connections. Represent each node and its connections accurately, "
    "ensuring all decision points and flows are included and connected. "
    "Now, generate BPMN business process model code in DOT language for the "
    "following textual description of a business process: "
)

description = (
    "The process begins when the customer submits an application. After submission, "
    "the application is reviewed by the credit officer. If the application is approved, "
    "the loan is disbursed. Otherwise, a rejection letter is sent. The process ends."
)

# Gemma 2 uses a single user turn without a separate system role
messages = [{"role": "user", "content": instruction + description}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=2048, temperature=0.1, top_p=1.0, do_sample=True)

dot_code = tokenizer.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
print(dot_code)

Training

Trained with LLaMA-Factory using LoRA on Gemma-2 9B Instruct. Detailed hyperparameters are reported in the PROFES 2025 paper. Training data: 21.5k cleaned instruction–input–output triples from MaD, split 80/10/10 for train/validation/test. The full splits are available at gcelikmasat-work/BPMN-IT-Dataset.

Limitations

• Scope. Control-flow slice of BPMN only (tasks, events, sequence flows, AND/XOR gateways). No pools, lanes, message flows, data objects, or sub-processes.
• Language. English only.
• Parameter efficiency. At 9B parameters, this model is roughly twice the size of the Qwen3-4B successor for comparable accuracy. For deployment-constrained settings, the 4B successor is preferred.
• Semantic equivalence. Structural similarity does not imply semantic equivalence, especially when input descriptions are ambiguous.

Citation

If you use this model, please cite the PROFES 2025 paper:

@inproceedings{celikmasat2025bpmg,
  title     = {Generating Business Process Models with Open Source Large Language Models using Instruction Tuning},
  author    = {{\c{C}}elikmasat, G{\"o}kberk and {\"O}zg{\"o}vde, Atay and Aydemir, Fatma Ba{\c{s}}ak},
  booktitle = {Product-Focused Software Process Improvement (PROFES 2025)},
  series    = {Lecture Notes in Computer Science},
  pages     = {269--284},
  year      = {2025},
  publisher = {Springer},
  doi       = {10.1007/978-3-032-12089-2_17},
  url       = {https://doi.org/10.1007/978-3-032-12089-2_17}
}

If you are comparing against this model as a baseline in a follow-up study, please also cite the journal extension:

@article{celikmasat2026instrubpm,
  title   = {InstruBPM: Instruction-Tuning Open-Weight Language Models for BPMN Model Generation},
  author  = {{\c{C}}elikmasat, G{\"o}kberk and {\"O}zg{\"o}vde, Atay and Aydemir, Fatma Ba{\c{s}}ak},
  journal = {Software and Systems Modeling},
  year    = {2026},
  note    = {Under review. arXiv:2512.12063},
  url     = {https://arxiv.org/abs/2512.12063}
}

Please also cite the source dataset:

@inproceedings{li2023mad,
  title     = {{MaD}: A Dataset for Interview-based {BPM} in Business Process Management},
  author    = {Li, Xiang and Ni, Lijuan and Li, Ran and Liu, Jiafei and Zhang, Ming},
  booktitle = {2023 International Joint Conference on Neural Networks (IJCNN)},
  pages     = {1--8},
  year      = {2023},
  publisher = {IEEE}
}

License

Released under the Gemma license, inherited from the base model (google/gemma-2-9b-it). Use is subject to Google's Gemma Prohibited Use Policy. The training data is distributed separately under the terms of the MaD dataset.