Automating code generation in manufacturing systems requires Artificial Intelligence (AI) models capable of interpreting textual requirement specifications. One of the main challenges is the absence of publicly available, domain-specific datasets suitable for training such models. This article presents AutoFactory, an open-source dataset that includes manually written and LLM-augmented requirement specifications, annotated by domain experts for Named Entity Recognition (NER) tasks using the BIO format.

Automating code generation in manufacturing systems requires Artificial Intelligence (AI) models capable of interpreting textual requirement specifications. One of the main challenges is the absence of publicly available, domain-specific datasets suitable for training such models. This article presents AutoFactory, an open-source dataset that includes manually written and LLM-augmented requirement specifications, annotated by domain experts for Named Entity Recognition (NER) tasks using the BIO format. AutoFactory enables the extraction and classification of key industrial components, including actuators, pre-actuators, sensors, effectors, and others. These elements correspond to real-world input and output signals in manufacturing systems. The dataset is grounded in realistic scenarios built with Factory I/O and includes 3D scenes, tag tables, and screenshots captured from multiple camera perspectives. To enhance linguistic diversity while preserving semantic intent, all specifications were expanded using large language models and evaluated through semantic similarity analysis (average cosine similarity above 0.92), complemented by manual validation to ensure consistency. To facilitate the annotation process, a custom labeling tool was developed. It runs locally to preserve data privacy and provides a user-friendly interface. The tool is adaptable to a wide range of NER tasks and allows researchers to customize labels for different domains by adding or modifying tags. To assess the practical value of the dataset, transformer-based models such as BERT and DistilBERT were fine-tuned on the annotated requirement specifications. The highest-performing model reached an F1-score of 0.95, confirming the dataset’s ability to support accurate identification of key components in manufacturing systems. This combination of human expertise and AI-based augmentation provides a robust foundation for training AI systems to interpret manufacturing behaviors from text. AutoFactory contributes to ongoing research efforts aimed at enabling intelligent systems to assist in generating control code from requirement specifications, reducing reliance on manual programming in industrial automation.

Modern Natural Language Processing relies heavily on supervised learning, which in turn depends on large volumes of carefully labelled text. Building such labelled datasets, particularly for tasks like Named Entity Recognition (NER), is one of the slowest and most expensive steps in any NLP project. Existing annotation tools fall short on three fronts: many depend on cloud services that cannot be used with confidential data, others demand heavy server installations that slow down small research teams, and several are locked behind commercial licenses. This paper presents AutoLabel-NER, an open-source, fully offline desktop tool that makes the creation of custom NER datasets faster, simpler, and safer. The software combines a clean graphical interface, linguistically accurate tokenization, intelligent label propagation across the entire document, and direct export to widely adopted formats compatible with modern NER training pipelines. By removing the need for an internet connection, paid licenses, or technical setup, AutoLabel-NER enables researchers in any domain to build high-quality annotated datasets in a fraction of the time required by conventional workflows.