bert-finetuned-bpmn

This model is a fine-tuned version of bert-base-cased on a dataset containing textual process descriptions.

The dataset contains 2 target labels:

• AGENT
• TASK

The dataset (and the notebook used for training) can be found on the following GitHub repo: https://github.com/jtlicardo/bert-finetuned-bpmn

Update: a model trained on 5 BPMN-specific labels can be found here: https://huggingface.co/jtlicardo/bpmn-information-extraction

The model achieves the following results on the evaluation set:

• Loss: 0.2656
• Precision: 0.7314
• Recall: 0.8366
• F1: 0.7805
• Accuracy: 0.8939

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	10	0.8437	0.1899	0.3203	0.2384	0.7005
No log	2.0	20	0.4967	0.5421	0.7582	0.6322	0.8417
No log	3.0	30	0.3403	0.6719	0.8431	0.7478	0.8867
No log	4.0	40	0.2821	0.6923	0.8235	0.7522	0.8903
No log	5.0	50	0.2656	0.7314	0.8366	0.7805	0.8939

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.0+cu116
• Datasets 2.7.1
• Tokenizers 0.13.2