Size matters less: how fine-tuned small LLMs excel in BPMN generation
Bibliographic record
Abstract
Abstract The generation of Business Process Model and Notation (BPMN) XML outputs from textual process descriptions presents a promising application for large language models (LLMs), yet it introduces significant challenges due to the structured and precise nature of process modeling. This study evaluates the performance of LLMs—Mistral, GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet—in BPMN generation, employing prompt engineering strategies across Simple, Medium, and Complex process descriptions to establish a baseline. Our findings reveal key limitations in LLMs, including limited output control, input presentation dependencies, and a lack of explainability, particularly for complex processes with nested flows and intricate dependencies. To address these challenges, we propose a novel Description-to-DOT pipeline utilizing a fine-tuned Qwen2.5 14B Coder model, trained on the MaD dataset of process description-DOT representation pairs. The novelty of the Description-to-DOT pipeline lies in its use of Graphviz DOT format as an intermediate representation, which requires generating fewer tokens and enables faster completion, followed by a Python script that converts DOT to BPMN XML in milliseconds—a significant efficiency improvement over the direct Description-to-BPMN pipeline, with the Description-to-DOT pipeline being approximately 6 times faster for Medium processes and 11 times faster for Complex processes. Experimental results demonstrate that the fine-tuned model significantly outperforms the evaluated LLMs, achieving accurate BPMN generation across all complexity levels. This study contributes: (1) Identification of LLM limitations in BPMN generation, such as logical inconsistencies, (2) A novel Description-to-DOT pipeline enhancing efficiency and accuracy, (3) A new benchmark dataset from the MaD dataset for Description-to-BPMN tasks, and (4) Comprehensive validation of the approach across complexity levels. These findings demonstrate the transformative potential of fine-tuned SLMs, with the Qwen2.5 Coder 14B enabling a scalable Description-to-DOT pipeline that excels in BPMN automation across complexity levels, validated on the MaD dataset.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.017 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.002 | 0.004 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.006 | 0.004 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".